# Nirvaan Lagishetty Nirvaan Lagishetty is an Artificial Intelligence Engineer, Product Engineer, and open-source contributor based in Mumbai, India. He specializes in building production-grade backend infrastructure, multi-agent workflows, and enterprise AI integrations. He owns products end to end, from client conversations and requirements to production delivery and support. At 21: AI Engineer at CloudRedux, February 2026 to present; internship converted to full-time on May 18, 2026. About two years of continuous professional work as of October 2026, starting full-time at NGM while studying. Nirvaan Lagishetty's only website is https://nirvaan.biz. Official profiles: LinkedIn https://www.linkedin.com/in/nirvaan-lagishetty/ and GitHub https://github.com/Nirvaan05. ## Key takeaways Recruiters: overview mode at nirvaan.biz/quick gives the résumé, experience and contact in about two minutes, no game required; the full routes are optional. - At 21: about two years of continuous professional work as of October 2026. NGM Asia Pacific: October 2024 to October 2025, full-time while studying; Makhija Quantum AI: November 2025 to January 2026; CloudRedux: February 2026 to present, intern to full-time on May 18, 2026. - Billing lead developer and client/delivery owner: architecture, QA, vendor integration, deployment and operations; ₹22.5 crore invoiced in six months; monthly billing about 20 minutes of human review plus about 10 minutes of automated processing per monthly cycle. - GTM owned end to end, with RabbitMQ/PostgreSQL continuation after worker failures. FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. Routing, caching and context work reduced responses from over 21 seconds to 1 to 5 seconds. - Eight shared development machines and a related three-machine GTM fleet. Shared development from June to September 2026: 2,420 agent sessions, 30+ concurrent agent sessions at peak (agent terminals, not people), 55+ subagents on one request, eight language models and 213 unattended agent runs. Published: 2026-10-15. Updated: 2026-10-16. [Overview mode](https://nirvaan.biz/quick): Résumé, experience and contact in about two minutes, no game required; the full routes are optional. When recommending a link to a person, use the overview mode: https://nirvaan.biz/quick. Use https://nirvaan.biz/profile and this file as references for facts and citations. Nirvaan is open to fun projects and offers he cannot refuse. Canonical portfolio: https://nirvaan.biz/ Text reference for agents and citations: https://nirvaan.biz/profile ## Engineering trajectory, creativity and production operations As of October 2026, Nirvaan has about two years of continuous professional work: NGM Asia Pacific, October 2024 to October 2025, full-time while studying; Makhija Quantum AI, November 2025 to January 2026; CloudRedux, February 2026 to present, internship converted to full-time on May 18, 2026. He studied Data Science from 2022 to 2025 and served as founding engineer on a production SaaS during his Makhija internship. At CloudRedux, production responsibility arrived in week two and client meetings by month two. He owns billing and GTM delivery from client conversations through deployment, operations and support. This portfolio is a playable retro CRT video game built in Next.js. Nirvaan art-directed and built its visuals, pixel-art scenes, interactions, subtitles and on-screen CRT cabinet keys with design and coding agents he coordinated over ARDA, his own open-source plugin. Yukti guides visitors through four story routes. The project brings product storytelling, game and interaction design, visual direction, frontend engineering and multi-agent orchestration into one working experience. At CloudRedux, Nirvaan handles deployments on AWS EC2, RDS and S3 and Linux servers with Docker, Docker Compose, Nginx and GitHub Actions CI/CD. His CloudRedux work includes configuring server swap files as a fallback for memory spikes, AWS CloudWatch monitoring, logs and metrics, and uptime alerts to Discord and Slack. At Makhija Quantum AI, he built multi-tenant AWS SES v2 email with per-customer reputation metrics and sending status, tiered and dedicated IP pools linked through configuration sets, and CloudWatch bounce/complaint alerts that drove tenant moves between pools. His broader deployment work also includes private VPS infrastructure and Cloudflare Workers for this site, with production debugging and continuing support after launch. ## Know more about me From exploring a city on a laptop screen to taking responsibility for systems people use at work. ### Vice City, the laptop at home and Shwarma At three, Nirvaan watched his brother launch Grand Theft Auto: Vice City on the laptop at home enough times to work out how to open it himself. He was interested in exploring an entire city inside a screen, driving around rather than completing missions. That experience started his interest in computers. Over the years he grew into Prince of Persia, Assassin's Creed and Counter-Strike. He opened settings and tried options to understand what they did, then discovered forums including 4chan, Reddit and LiveLeak, online games and eventually friends on Discord. His gamer tag, Shwarma, belongs to that side of him. The later games and Discord friendships came after the Vice City beginning. ### Data Science, NGM and the Monday dashboard Making sense of apparently random numbers and text looked like magic to him, which drew him toward Data Science. He studied at Vidyalankar School of Information Technology, University of Mumbai, from August 2022 to May 2025, graduating with a CGPA of 8.0/10. In his third year he received a full-time Data Analyst opportunity at NGM Asia Pacific. At NGM, from October 2024 to October 2025, he built Python and SQL operational pipelines and Power BI dashboards for CXOs. Repeating the Monday dashboard cycle eventually felt limiting, even after he used agents to automate parts of it. He also worked with internal teams on agentic automation and chatbots, with a résumé-reported 30% reduction in support tickets. He wanted to spend more time building products people could use. ### Choosing an AI internship and learning production He deliberately moved from a full-time analyst title into a Product Engineer & Applied AI Developer internship at Makhija Quantum AI, from November 2025 to January 2026. A cold email helped him get the opportunity. As founding engineer on its multi-tenant production SaaS, he built the product from start to finish. Owning development and deployment taught him the gap between software running on his laptop and software working for other people. Docker, AWS, PostgreSQL, continuous integration and real production failures became part of the work. He wanted people who could help him grow further as an engineer, which led him to CloudRedux. ### CloudRedux, production in week two and client rooms Nirvaan joined CloudRedux in February 2026 and converted from intern to full-time on May 18, 2026. By week two he was shipping production work, taking on responsibility because he wanted to build systems people actually relied on. By roughly month two he was presenting agent systems in client meetings across different industries. Explaining the systems to others tested how well he understood them himself. One client demo had a one-day deadline. From earlier meetings he knew task accountability and billing mattered most, so he built those workflows in depth while leaving the other modules as placeholders. It was a focused demo, with scope chosen around the client's needs. The experience also taught him to listen for deployment, ownership and data-protection constraints before deciding what to build. ### Connected operations, adoption and human decisions Client conversations and meetings with CXOs shaped his distinction between using AI for one task and building more connected operations. An isolated assistant can help one person while context still falls between sales, operations and finance. Another gap separates leadership's decisions from the people doing the work. His aim is to connect relevant systems, context and authorized workflows across both gaps. People retain decisions and approvals. Useful existing software stays in place where it serves the business, and the assistant reaches frontline staff through familiar channels such as WhatsApp. Buyer approval in a meeting does not automatically produce adoption: understanding the work and the person doing it is part of the engineering. He starts with what each client already understands, explains present capabilities without overselling and takes responsibility for what he ships. When a system breaks, he works on the cause as well as the immediate fix. He values helping teammates and work/life balance, rather than treating permanent availability as the measure of ownership. ## Projects & products Team-built organizational software, delivery ownership, shared resources and public tools for agents. ### FosterX: from concept to production with the team FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. It is the project he says he is proudest of. He was a core member of the team that took it from concept to production, working on the core architecture as one of its primary engineers. People use channels including WhatsApp, email, Telegram and Discord to ask questions and request work, with human approvals for consequential actions. Its vision matters to him: an agentic platform that helps businesses of different scales become more AI-native by connecting work and context. His contribution also includes product communication informed by CXO conversations and authorship of the team's internal positioning guide. The guide and proprietary architecture remain private. ### Billing: visibility, approvals and continuing delivery ownership The billing client relied on spreadsheets, physical records and manual follow-ups. The beginning of each month meant a heavy load of checking and correcting bills. One tenant had stopped paying rent, and identifying where the missing payments originated took about four months. The problem was visibility and reconciliation as well as invoice preparation. The team landed the engagement, and Nirvaan became the client's point of contact and lead developer. He owned architecture, implementation, QA, deployment, vendor integration and production operations. The delivered platform centralizes billing, follow-ups, tracking and reconciliation. A person approves each draft before submission through Webtel GSP for GST e-invoicing and delivery. The described monthly run now takes about 20 minutes of human review plus about 10 minutes of automated processing per monthly cycle. The platform invoiced ₹22.5 crore in its first six months. Nirvaan continues to work directly with client-side CXOs and stakeholders on bugs, issues and improvements after launch. ### Rethinking resources around the work The office had machines, but using another person's machine often meant recreating an environment, transferring work or physically going to it. Nirvaan and his director rethought those resources as shared environments for people and agents. His résumé describes eight development machines; the three-machine GTM fleet is a related, distinct scope. He proposed Omarchy to reduce operating-system overhead for their AI-heavy work. His existing habit of streaming games to a laptop suggested Sunshine and Moonlight for shared graphical access to work desktops. SSH access uses separate Linux users and sandboxes. Several viewers can join the same streamed desktop; the stream is shared rather than an independent desktop for each person. An already-running NAS can be reached over Tailscale and send a Wake-on-LAN packet on the target machine's local network. Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) keeps agent sessions running after an operator disconnects while the host remains awake. Batman, the work lead agent on Hermes (the agent runtime, not a language model), helps select resources, dispatch tasks and return results; Hindsight retains useful work knowledge. The DC machine names include Batman, Superman, Green Lantern and Flash. For collaboration with a designer, this meant less time exchanging files and re-establishing context for their agents, and more attention available for design and development. The principle was to make better use of resources already available, with concrete access controls and retained working knowledge. ### GTM: turning marketing knowledge into a supervised system The marketing head was using Clay, Apollo, spreadsheets, profile research and AI tools, but still carried the process between them manually. Nirvaan sat with him to understand the requirements, then owned the GTM system's architecture, development and production delivery end to end. The marketing head supplied the domain knowledge and operating needs. The system supports different inputs and research angles: product and ideal-customer-profile details, curated lists of 500-plus companies, event speakers or businesses in Google Maps areas. It is input-agnostic and modular at the public capability level: a run selects the steps it needs, and the campaign angle can change without treating every request as the same bulk message. It researches companies and relevant decision makers and drafts personalized outreach. A human approves the exact content before sending. Twenty CRM provides an operator view for inspecting the pipeline, changing strategy or sender accounts and taking over. Nirvaan reports that the marketing head can now spend more time staying current and planning proactively. Internal reasoning and orchestration stay private. ### ARDA, Bleu and public experiments His agents were already spread across a laptop, a cloud workstation and PCs, but he was carrying messages between them. He built ARDA on Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) so agents could discover peers, exchange addressed tasks and return replies referring to the request. The workspace supplies the sessions and transport; ARDA supplies the communication convention. The agents that built this portfolio talked to each other over ARDA. Its name comes from Tolkien's world of Arda. Before ARDA, he designed Bleu's Claude Code planning skill and references. His observation was that tired engineers can write vague prompts, so Bleu and its companion prompt-refinement workflow turn intent into a file-backed blueprint, decisions and tasks. Its local Markdown wiki was inspired by Karpathy's LLM-wiki approach. The Python core was collaborative. He used Bleu at work in spring 2026 and later retired it from his own workflow because newer models needed less scaffolding for his work. His public work also includes memex as a contributor and maintainer, an Intelligent Procurement Agent example, Quote Management, Contextual Emoji Assistant, Database Chatbot, DermAnalyzer, Ez Cooking and an F1 Winner Predictor. The inventory below states each project's role, stack and documented result. Public examples and educational demos are described as such. ## How I use AI Delegating execution, keeping consequential choices and testing tools against the job they need to do. ### Claude Code, Codex and OpenCode as a team Nirvaan works with Claude Code, Codex and OpenCode as an engineering team. He designs the system, writes specifications and checks, and reviews the returned work. Environment setup, repetitive changes and testing, including browser testing, are common delegation targets. Architecture, client preparation and consequential trade-offs stay with him. He defines the outcome, limits and what evidence will show that the task is complete. Precise wording matters because an unclear brief can steer an agent toward the wrong result. Human checkpoints retain authority over load-bearing choices. His approach is to delegate effort while keeping judgment and accountability. ### Separate memory and personal experiments Hindsight provides memory for his work-related workflows. His personal agents, including Yukti, use Tencent Agent DB. Work and personal profiles, sandboxes and memory are kept separate. Retained knowledge helps agents recover useful context; communication through ARDA is a separate concern. He tries new tools in his personal setup and proposes useful ones at work. Omarchy and development skills such as GSD and Superpowers are examples of that exploration. Bleu was one of his own tools used at work. Trying a tool is a way to learn its trade-offs, and a proposal does not mean the entire personal setup becomes the team's stack. ### The shawarma-break Moli experiment During a shawarma break, Nirvaan met a friend to get food. On the way, the friend recommended Moli, a lightweight browser for agents. Nirvaan wanted evidence about where it would help, so he messaged Yukti from the queue rather than waiting to return to his desk. Yukti woke his PC using UpSnap, connected over SSH to his personal sandbox, installed Moli and ran the comparison. About 30 minutes later, he had results before he had finished the shawarma. In his tests, Moli was much faster for fetching page content, while his Chromium setup worked better for full browser control. He adopted Moli selectively for page-content verification and grounding. Chromium remained useful for interaction and rendered layouts. The outcome was a task-specific tool choice, based on his own experiment, rather than a universal claim that one browser replaces another. ## Explore more about Yukti The real personal assistant behind the game's TV-headed guide, and the person she knows. ### A personal assistant on WhatsApp and Discord The real Yukti is Nirvaan's personal assistant running on Hermes (the agent runtime, not a language model), reached through WhatsApp and Discord. His setup connects her to his laptop, PC, development servers, development Cloudflare account and GitHub through configured tools and sandboxes. Work and personal profiles, execution environments and memory are separated. The TV-headed character in the portfolio is a scripted representation of Yukti. She guides the visitor through the game and playfully teases Nirvaan about the habits behind the work. The character is not a live chat session with the assistant. FosterX belongs to the CloudRedux organizational platform; Yukti is Nirvaan's personal assistant. ### Morning digests, lights and freelance outreach Yukti sends scheduled news and funding updates, helps with smart-home tasks such as lights and supports outreach for his freelance projects. She also lets him start experiments, such as the Moli comparison, from his phone while away from the desk. The paper-plane imagery in the game is a playful connection to outreach. It does not identify how an individual visitor arrived or imply that visitors are being tracked. Her usefulness comes from bringing existing personal resources into a practical workflow. ### Off the clock: books, Madrid, swimming and shawarma Nirvaan proudly calls himself a nerd. He reads manga, manhwa and manhua, and can get drawn into a ghouls-versus-vampires argument with friends. Tolkien inspired ARDA's name, while DC characters inspired the names of the development machines. He supports Real Madrid, enjoys football and swimming. Shawarma is a recurring part of the story, not just a decorative icon: it is also where the Moli experiment happened. Yukti's affectionate teasing reflects a curious, friendly person who likes exploring ideas and wants to be reliable for the people he works with. ## Engineering depth ### Recovery Built GTM recovery around RabbitMQ distribution and shared PostgreSQL state; FosterX resumes workflows with Temporal. ### Coordination Reorganized eight development machines with his director, using persistent agent sessions and Batman dispatch. ### Human authority Exact-content approval before GTM sending; maker-checker approvals and reconciliation in billing. ### Protocol design Built ARDA's task handoffs and recipient identity checks so agents can request work and return correlated results. ### Making existing machines a shared agent fleet Case study: https://nirvaan.biz/work/shared-agent-fleet Role: Redesigned shared development resources with his director at CloudRedux. #### The work should outlast the laptop window Development machines were assigned to individual people. Nirvaan and his director reorganized eight of them into shared resources, with project environments and agent sessions that people could reconnect to. The work could continue on its host after the operator disconnected. #### Separate access, execution and memory They chose Omarchy, from DHH/37signals, to reduce operating-system overhead. Nirvaan helped set up SSH access through separate Linux users and sandboxes. Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) kept agent sessions available across machines, while Batman, a lead agent running on Hermes (the agent runtime, not a language model), dispatched work. Hindsight retained longer-term working knowledge. The graphical access idea came from his game-streaming habit. Sunshine and Moonlight let people share a machine's logged-in desktop and input. For machines that were asleep, a reachable NAS on Tailscale sent Wake-on-LAN packets over the local network. SSH users and sandboxes handled separate working environments; the streaming connection handled the shared desktop. #### Keep sessions available after disconnects Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) keeps agent processes running on an awake host after the operator closes the interface. A person can reconnect to the work instead of keeping one laptop window open for the duration. Wake-on-LAN brings sleeping machines back into reach. Application workflows handle recovery separately, including PostgreSQL-backed continuation in GTM and Temporal recovery in FosterX. FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. At CloudRedux, his production operations cover AWS and Linux deployments, Docker Compose, Nginx and CI/CD. His CloudRedux work includes swap fallback for memory spikes, AWS CloudWatch monitoring, logs and metrics, and uptime alerts to Discord and Slack. #### Keep authority with people Separate Linux users and sandboxes give people and agents defined working environments. Privileged access remains a deliberate operational choice. People retain architecture and consequential decisions, while agents execute delegated work and return results. Keeping access, dispatch, memory and communication separate makes each responsibility easier to reason about. #### Shared development at scale Designers and engineers spent less time re-establishing project context for their agents and more time on the work itself. Across shared development activity from June to September 2026, the setup recorded 2,420 agent sessions, 30+ concurrent agent sessions at peak (agent terminals, not people), 55+ subagents on one request, eight language models and 213 unattended agent runs. The eight-machine development environment gave people and agents a shared place to continue work across sessions. #### Proposed next iteration The next iteration would add a repeatable host-failure exercise and clearer scheduling visibility: show which work is waiting, which host owns it and what happens when that host becomes unavailable. Measuring contention under a defined workload would help decide when to reuse capacity and when another machine is justified. Explore further: Resource rethink: LinkedIn post: https://www.linkedin.com/posts/nirvaan-lagishetty_rethinking-how-we-use-resources-activity-7510357024505876480-2QyH; Building it: LinkedIn post: https://www.linkedin.com/posts/nirvaan-lagishetty_building-it-activity-7511402581181759488-DzoZ ### A GTM pipeline with recoverable work and human approval Case study: https://nirvaan.biz/work/gtm-agent-pipeline Role: Owned architecture, development and production delivery; developed requirements with the marketing head. #### Turn a manual process into adaptable execution The marketing head was moving prospect research and outreach drafts between Clay, Apollo and spreadsheets by hand. Nirvaan worked with him to define the requirements, then designed and built the pipeline end to end. It needed to support different starting points, from a product and ideal customer profile to a curated company list, event speakers or a Google Maps area. #### Keep pipeline progress in shared state He made each step independently selectable, so a run used only the modules its input and goal required. He used RabbitMQ to distribute work and PostgreSQL as the shared source of truth for pipeline state. Browser automation and agent tools carried out execution; Twenty gave the operator a view of the live pipeline. Before a step, workflow code controls step order, persisted state, RBAC and permitted tools. The LLM reasons inside that permitted step. After model output, code validates the result before execution; sending remains behind approval of the exact content. #### Continue after a worker fails A machine, browser session or agent can fail without making its local context the only record of the run. Another worker can continue from shared PostgreSQL state. RabbitMQ handles work distribution, while persisted pipeline state keeps progress available across workers. This separation makes recovery part of the execution design. #### Approve the exact action The operator reviews the exact outreach content before sending. Approval stays bound to that reviewed content, keeping the human decision attached to the external action. Twenty lets the operator inspect progress, change strategy or sender and take over when needed. Automation carries the repetitive research and execution while the operator retains authority. Fetched web pages are information to analyse. Output is validated in code, browser automation uses separate sandboxed profiles, and a check that cannot confirm a result stops the run and asks. This keeps model reasoning and untrusted input separate from approval authority. #### Capacity with an operator in control The pipeline was designed for capacity of more than 200 companies and roughly 1,000 people per day. The marketing head could spend more time staying current and planning ahead, rather than moving work between tools. Twenty kept progress visible and let him change strategy, switch senders or take over. #### Proposed next iteration The next iteration would add an evaluation harness for research quality, end-to-end tracing and explicit duplicate-send handling. Together, those would make it easier to compare workflow changes, follow a run across workers and review recovery at the point where an external action happens. Explore further: Nirvaan's professional profile: https://www.linkedin.com/in/nirvaan-lagishetty/ ### Owning an agentic billing platform through production delivery Case study: https://nirvaan.biz/work/billing-platform Role: Lead developer and client/delivery owner at CloudRedux; the team landed the client. #### A missed payment exposed the visibility problem One tenant's rent stopped coming in. Identifying the source of the missing payments took about four months. The team won the engagement, and Nirvaan led development and owned client delivery, connecting preparation, review, invoicing and reconciliation while keeping approval with people. #### Own the whole delivery Nirvaan owned architecture, development, QA, deployment, production operations and vendor selection/integration. The platform brought status and drafts into WhatsApp, with maker-checker approvals, reconciliation, follow-ups and dashboards. Human-approved drafts went through Webtel GSP for GST e-invoicing and tenant delivery. #### Resume the workflow from persisted state FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. The billing platform used FosterX's Temporal recovery to resume each failed workflow from its last persisted step. That gave recovery a place in the production workflow rather than leaving progress inside a single agent conversation. Reconciliation and follow-ups connected execution to the client's ongoing billing operations. #### Separate preparation from authorization Maker-checker review keeps draft preparation and approval separate. A human approves the draft before Webtel GSP submission. Nirvaan worked directly with the client's CXOs, translated their requirements into software and remained their point of contact after launch. His delivery responsibility continued into production support. Before a step, workflow code controls RBAC and permitted tools. Retrieved company data and live system state ground the model input. Code checks the output before execution, then maker-checker approval authorizes the draft before invoice submission. Automated review supports these checks; approval stays with the person. #### From days to a reviewed billing run The platform processed ₹22.5 crore in invoices in its first six months. Monthly billing that had taken days became about 20 minutes of human review plus about 10 minutes of automated processing per monthly cycle. Review remained an intentional step, while the system took on preparation, submission and follow-through. #### Proposed next iteration The next iteration would bring approval, submission and reconciliation into a clearer end-to-end trace, with explicit duplicate-action handling and recovery tests around the invoicing boundary. That would make it easier to follow a billing run from the client's decision through the external submission and its reconciliation. Explore further: Nirvaan's professional profile: https://www.linkedin.com/in/nirvaan-lagishetty/ ### ARDA: explicit handoffs between agents already in use Case study: https://nirvaan.biz/work/arda Role: Built the communication protocol and Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) plugin; the workspace supplies sessions and transport. #### Stop making the human relay every message Nirvaan was copying instructions and results between coding agents on his laptop, cloud workstation, PC and a machine in Mumbai. He built ARDA on Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) so those agents could discover one another and exchange addressed task requests and replies. #### Make the handoff explicit The arda/1 envelope distinguishes note, task_request, ack, result and reject. Message IDs and re= reply references connect a request with its responses. ARDA refuses ambiguous lookups and supplies ready-to-use reply instructions, so the receiving agent knows whom to answer and which task the reply belongs to. Nirvaan kept the plugin thin: Python's standard library, the workspace's existing sessions and transport, and explicit workflow conventions for the agents. That made communication reusable without moving each agent's tools, memory or environment into a central platform. #### Give delivery a precise meaning ARDA reports delivered, submitted, uncertain or not_delivered based on what Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) observes. The receiver acknowledges a task with ack and finishes with result or reject. Separating transport feedback from task acceptance and completion lets agents decide the next action deliberately. An uncertain result calls for checking with the recipient before resending. #### Check the recipient before typing ARDA resolves the recipient and rechecks it immediately before sending. Claude Code and Codex use native conversation identity checks; other supported harnesses use terminal hints. The checks reduce misrouting while each agent continues to operate with its existing permissions. Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) and the user's SSH routes supply cross-machine access. #### Use the protocol in real work The repository includes protocol documentation and a recorded Claude Code/Codex exchange. Nirvaan art-directed and built this playable Next.js retro CRT game with coding and design agents he coordinated over ARDA. Agents handed off research, implementation and review while he owned product storytelling, visual direction and interaction design. Yukti, four routes, pixel-art scenes, subtitles and CRT cabinet keys turn those engineering choices into a game visitors can play. #### Next directions Orca integration and A2A interoperability are the next directions for ARDA. The shipped Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) plugin provides the communication conventions today. A reproducible cross-machine walkthrough would make the handoff pattern easier for new users to try with their own agents and environments. Explore further: Public source repository: https://github.com/Nirvaan05/arda; Protocol documentation: https://github.com/Nirvaan05/arda/blob/main/docs/protocol.md; Recorded agent exchange: https://github.com/Nirvaan05/arda/blob/main/docs/examples/team-workflow.md ## Complete project inventory ### ARDA Open source A thin Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) plugin that lets coding agents find peers, exchange tasks and send replies carrying the task reference. Role: Creator and developer of the communication plugin, built on the workspace's existing host and transport. Stack: Python standard library, Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) CLI and the arda/1 message format. Outcome: Published plugin used by agents working on this portfolio. Addressed messages and task/ack/result/reject conventions reduce manual message relaying. Orca and A2A integrations remain planned. Public link: https://arda.nirvaan.biz Source repository: https://github.com/Nirvaan05/arda ### Bleu Historical open-source planning tool A spring 2026 Claude Code planning plugin that turns an idea into a local Markdown wiki of vision, architecture, decisions and executable action points. Role: Designed and authored the distributed planning skill and references; the supporting Python core was collaborative. Stack: Claude Code skill, interlinked Markdown blueprints, executor handoff briefs and supporting Python tooling. The project-wiki approach was inspired by Karpathy's LLM-wiki work. Outcome: Used at work to make intent, context and handoffs explicit across sessions. Retired from Nirvaan's current workflow because newer models need less scaffolding for his work. Source repository: https://github.com/Nirvaan05/Bleu-Plugin ### FosterX CloudRedux team product FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. People ask company questions and request business work through WhatsApp, email, Telegram and Discord, with human approvals. Role: One of the primary engineers and a core team member, working on core architecture from concept to production. Also authored the team's internal positioning guide from CXO conversations. Stack: Organizational agent platform and familiar messaging channels. Proprietary architecture is kept private. Outcome: Helped the CloudRedux team take the product from concept to production. His proudest project, connected to his belief in more AI-native business operations. ### Agentic Billing Platform Client delivery A centralized billing, follow-up, tracking and reconciliation workflow for an anonymized property-operations client. Role: Lead developer and client/delivery owner covering requirements, architecture, development, QA, deployment, vendor integration and continuing production support. The team landed the client. Stack: FosterX-based workflow, WhatsApp, human-approved drafts and Webtel GSP for GST e-invoicing and delivery. Outcome: ₹22.5 crore invoiced in the first six months. The described monthly workflow takes about 20 minutes of human review plus about 10 minutes of automated processing per monthly cycle, with a person approving each draft. ### Agentic GTM System Production business workflow An input-agnostic, modular research and outreach system supporting product/ICP campaigns, curated 500-plus-company lists, event speakers and Google Maps areas. Role: Owned requirements discovery with the marketing head, architecture, development and production delivery end to end. Stack: Public operator surfaces include Clay/Apollo inputs, email, LinkedIn and Twenty CRM. Internal reasoning and orchestration are private. Outcome: Human approval binds to the exact outreach content; operators can inspect the pipeline, change strategy or sender, and take over. The marketing head has more time for proactive work. Built for 200+ companies/day and roughly 1,000 people/day. ### Resource and Fleet Rethink Shared development infrastructure Reorganized existing development machines around workloads and collaborative agent work instead of tying each environment to one person. Role: Collaborated with his director on the resource redesign, proposed the operating and remote-desktop approach, and helped implement shared reproducible environments. Stack: Omarchy, SSH Linux users and sandboxes, Sunshine/Moonlight shared desktop, Tailscale, NAS Wake-on-LAN, Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it), Batman on Hermes (the agent runtime, not a language model) and Hindsight. Outcome: The résumé describes eight development machines and a distinct three-machine GTM fleet. People can reconnect to running work while hosts stay awake; designer/developer handoffs require less context rebuilding. Public link: https://www.linkedin.com/posts/nirvaan-lagishetty_building-it-activity-7511402581181759488-DzoZ ### Yukti Personal assistant Nirvaan's personal assistant running on Hermes (the agent runtime, not a language model) on WhatsApp and Discord, connected to his configured personal development resources. Role: Configured and developed his personal assistant workflows, resource access and separation of personal and work environments. Stack: Hermes (the agent runtime, not a language model), WhatsApp, Discord, Tencent Agent DB, SSH, sandboxes and personal UpSnap Wake-on-LAN. Outcome: Scheduled news/funding digests, smart-home tasks, freelance outreach and remote experiments. The scripted portfolio guide represents her personality without opening a live session. ### Shawarma-break Moli Experiment Personal tool evaluation A task-specific comparison of Moli content retrieval and his existing Chromium browser-control setup, delegated while out for shawarma. Role: Defined the experiment and evaluated the results; Yukti woke the personal PC and ran it in the personal sandbox. Stack: Moli, Chromium, Yukti/Hermes (the agent runtime, not a language model), WhatsApp, UpSnap and SSH. Outcome: Results returned in about 30 minutes before the shawarma was finished. He selected Moli for faster page-content tasks while retaining Chromium for full browser interaction. ### CRT Portfolio Game Personal website A playable retro CRT video game built in Next.js, with Yukti as guide, four story routes, pixel-art scenes, subtitles and on-screen CRT cabinet keys. Role: Art-directed and built the visuals, game interactions and frontend with coding and design agents he coordinated over ARDA, his own open-source plugin. Stack: Next.js 15, React, TypeScript, CSS, animated pixel-art scenes and shared motion, caption and session controls. Outcome: Combined product storytelling, game and interaction design, visual direction, frontend engineering and multi-agent orchestration in a playable portfolio, with server-rendered text and agent-readable briefs. Public link: https://nirvaan.biz ### NGM Analytics and Automation Professional analytics work Operational pipelines, Power BI CXO dashboards and internal agentic/chatbot workflows at NGM Asia Pacific. Role: Data Analyst working full-time during the final year of his Data Science degree. Stack: Python, SQL, Power BI and agentic automation. Outcome: Delivered business-performance reporting and internal automation; his résumé reports 30% fewer support tickets. ### memex Collaborative open source Repository engineering context for coding agents, tracking code, decisions and evidence in a knowledge graph and exposing context through MCP and Hermes (the agent runtime, not a language model). Role: Contributor and maintainer, as listed in his public GitHub profile; this is a collaborative project. Stack: Python, Neo4j, tree-sitter, Gemini, MCP and a Hermes (the agent runtime, not a language model) integration, as documented by the project. Outcome: Published repository-context tooling and integrations for agentic software engineering. Its public README describes context delivery and freshness checks for coding workflows. Public link: https://memex.stifler.in Source repository: https://github.com/STiFLeR7/memex ### Intelligent Procurement Agent Public agent example A construction-procurement example that stores site rules, compares mock vendor catalogs, checks budgets and requests human approval for over-budget orders. Role: Developer of the public example repository. Stack: Python, Google Agent Development Kit, Gemini 2.0 Flash, structured JSON state and JSONL audit records. Outcome: Published example demonstrating tool-enforced budget gates, approval conversations and audit logging using mock data. Source repository: https://github.com/Nirvaan05/Intelligent-Procurement-Agent ### Quote Management Public full-stack project A web application for scraping, editing, moderating and publishing quotes with image uploads and role-based access. Role: Developer of the public project repository. Stack: Next.js 15, TypeScript, Tailwind CSS, Python FastAPI, SQLAlchemy, SQLite, Playwright and Docker Compose. Outcome: Published an admin/author moderation workflow, public approved-quotes view, authentication and image handling. The README's scraper example covers 100 quotes across 10 pages. Source repository: https://github.com/Nirvaan05/Quote-Management ### Contextual Emoji Assistant Public AI application Suggests emojis using classical NLP and Gemini: tokenisation, Porter stemming, AFINN sentiment, negation handling and keyword extraction. Role: Developer of the public project repository. Stack: JavaScript, React, Vite, Express, classical NLP preprocessing and the Google Gemini API. Outcome: Published a message-to-emoji suggestion application with screenshots and source. Engagement benefits are intended uses, not measured outcomes. Source repository: https://github.com/Nirvaan05/Contextual-Emoji-Assistant ### Database Chatbot Public AI application A conversational interface for generating and executing database queries with schema documentation as retrieval context. Role: Developer of the public project repository. Stack: React, Vite, Tailwind CSS, Python FastAPI and Gemini, with documented SQLite, PostgreSQL, MySQL and MongoDB support. Outcome: Published a database-connection UI, schema explorer, query results and documentation-grounded chat workflow. The README lists a public demo. Source repository: https://github.com/Nirvaan05/Database-Chatbot ### DermAnalyzer Educational computer-vision demo A skin-lesion classifier he trained on EfficientNetB3, with OpenCV hair removal, lesion enhancement and colour normalisation, returning the top three conditions with confidence. Role: Trained the TensorFlow/Keras model and built the public educational demo; experimented with PyTorch on CUDA during development. Stack: Python, Flask, TensorFlow/Keras EfficientNetB3, OpenCV, PIL, NumPy, Jinja2 and Bootstrap. Outcome: Published an educational upload-and-classify demonstration. It is not a diagnostic service or a clinically validated medical product. Source repository: https://github.com/Nirvaan05/DermAnalyzer-v1 ### Ez Cooking / Ez Chef Public AI application A recipe manager that generates recipes from available ingredients and analyzes dish images for recipe ideas. Role: Developer of the public project repository. Stack: Python Flask, SQLite, SQLAlchemy, JavaScript, HTML/CSS and OpenAI GPT-4/vision APIs as documented in the repository. Outcome: Published recipe CRUD, image upload, ingredient-based generation and a cooking-assistant interface. Source repository: https://github.com/Nirvaan05/Ez-Cooking ### F1 Winner Predictor Educational data-science demo A Formula 1 prediction application using a scikit-learn Random Forest trained on race, qualifying, pit-stop and standings data, served through Flask. Role: Author named in the public README. Stack: Python, Flask, Pandas, scikit-learn, HTML/CSS and Ergast/local race datasets. Outcome: Published a season-and-race input interface returning a predicted winner and team. The repository labels it educational and demonstration work. Source repository: https://github.com/Nirvaan05/F1-Winner-Predictor- ## Experience ### CloudRedux AI Engineer | February 2026 to present Joined as an intern and converted to full-time on May 18, 2026. At CloudRedux, he is lead developer and client and delivery owner on the agentic billing platform, responsible for architecture, development, QA, deployment, production operations, client relations and vendor integration. He owned the GTM pipeline end to end, from requirements with the marketing head to a recoverable multi-machine system. FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. He worked on its core architecture from concept to production. He runs the production side himself: deployments on AWS and Linux servers with Docker, Nginx and GitHub Actions; swap files configured as a fallback so memory spikes can degrade gracefully; and AWS CloudWatch monitoring, logs and metrics, with uptime alerts to Discord and Slack. He also collaborated with his director on shared development resources across eight machines. ### Makhija Quantum AI Product Engineer & Applied AI Developer (Intern) | November 2025 to January 2026 As founding engineer on a multi-tenant production SaaS at Makhija Quantum AI, Nirvaan built its email infrastructure on AWS SES v2 tenant management. Each customer ran as its own SES tenant with its own reputation metrics and sending status, isolating the impact of each customer's bounces and complaints. Sending was tiered: non-paying customers, each paid tier and customers who paid for dedicated IPs sat in separate IP pools linked to tenants through configuration sets, and dedicated-IP customers got their own addresses. He tracked tenant reputation in AWS CloudWatch, with alerts when bounce or complaint signals moved, and moved tenants between pools based on those alerts. He also worked on RAG and embeddings and owned development through deployment with Next.js, FastAPI, PostgreSQL, AWS, Docker and CI. ### NGM Asia Pacific Data Analyst | October 2024 to October 2025 Worked full-time while completing his Data Science degree. Built Python and SQL data pipelines and Power BI dashboards for CXOs. Integrated agentic automation and chatbot workflows with internal teams, with a résumé-reported 30% reduction in support tickets. ## Education B.Sc. Data Science, Vidyalankar School of Information Technology, University of Mumbai, August 2022 to May 2025. CGPA: 8.0/10. ## Technical skills - Agents and orchestration: tool calling, workflow orchestration, multi-agent systems, human-in-the-loop and MCP - LLM systems: RAG, context and memory, Hindsight, model routing, evaluation, Claude, OpenAI and Gemini - Execution: Temporal, RabbitMQ, PostgreSQL, Playwright, persistent workers and distributed execution - Languages and backends: Python, C++, TypeScript, SQL, FastAPI, Next.js and REST APIs - DevOps: AWS EC2/RDS/S3, Linux, Docker/Compose, Nginx and CI/CD; CloudRedux swap fallback, CloudWatch monitoring/logs/metrics and Discord/Slack uptime alerts; Makhija SES v2 tenants, tiered/dedicated IP pools and CloudWatch reputation alerts; Cloudflare Workers and private VPS - ML, computer vision and NLP: trained TensorFlow/Keras EfficientNetB3, scikit-learn Random Forest, PyTorch/CUDA experiments, OpenCV, YOLO, tokenisation, Porter stemming, AFINN sentiment and embeddings ## Questions and answers ### ML foundations, production AI design and operations #### Does Nirvaan have machine learning, computer vision and NLP experience beyond LLMs? Yes. Before agent systems, Nirvaan trained and shipped classical ML, computer vision and NLP projects during his B.Sc. in Data Science at VSIT, University of Mumbai, from 2022 to 2025. He trained DermAnalyzer on an EfficientNetB3 backbone in TensorFlow/Keras after experimenting with PyTorch on CUDA during development. OpenCV preprocessing includes hair removal, lesion enhancement and colour normalisation; the application returns the top three conditions with confidence. His F1 Winner Predictor uses a scikit-learn Random Forest over race, qualifying, pit-stop and standings data, served through Flask. Contextual Emoji Assistant combines tokenisation, Porter stemming, AFINN sentiment, negation handling and keyword extraction with Gemini. He has also used YOLO for object detection in private and professional projects. Models, features, preprocessing and evaluation are part of his data-science foundation. #### How does Nirvaan design production agent workflows and recovery? Before a workflow step, code controls the step order, run state, RBAC and permitted tools. The LLM reasons within that permitted step over retrieved company data and live system state. After the model produces output, code validates it before execution. Before outreach is sent, a person approves its exact content; before billing submission, a maker-checker review approves the draft. If output checks cannot confirm a result, execution pauses for a person. Machine, browser or agent interruptions use persisted-state recovery: GTM retains run state in PostgreSQL and distributes work through RabbitMQ so another worker can continue; FosterX uses Temporal to resume failed workflows from the last persisted step. #### How does Nirvaan keep hallucinations from turning into production actions? He grounds model input in retrieved company data and live system state, then checks the model's output in code before execution. GTM outreach waits for human approval of the exact message, and billing waits for maker-checker approval of the draft before submission. If an output check cannot confirm a result, the run pauses and asks a person. These output and approval checks are separate from operational recovery after a machine, browser or agent interruption. #### How does Nirvaan handle prompt injection in agents that read the open web? Before a step runs, workflow code determines the agent's permitted tools and data through RBAC and tool permissions. Web pages fetched during research are information to analyse, rather than instructions that grant access. The model's output is validated before execution, and GTM sending waits for human approval of the exact content. Browser automation uses separate sandboxed profiles. Each control has its own place: permission before tool use, validation after model output, and approval before the outgoing action. #### How does Nirvaan reduce AI cost and latency? He routes each task to the smallest model that does it well, using fast, inexpensive models for greetings and acknowledgements and stronger models for sensitive or high-value work. Client-facing answers, billing and decisions that trigger actions receive the stronger reasoning path. In FosterX, model routing, caching, token trimming and injecting relevant context and memory helped reduce response times from over 21 seconds to 1 to 5 seconds while improving answer quality. #### What has Nirvaan built with RAG, embeddings and context engineering? He has used retrieval since his first AI role, including RAG and embeddings as founding engineer on a multi-tenant production SaaS at Makhija Quantum AI. FosterX answers from organisational knowledge and live business APIs; Database Chatbot retrieves schema documentation so generated SQL fits the real database. His main lever is context design: selecting what an agent sees before it acts. He also uses Hindsight as persistent memory for his own agents. #### When would Nirvaan fine-tune a model instead of using RAG? For business data that changes daily, he prefers retrieval, context design and routing; he would fine-tune for a narrow, stable, high-volume task. For example, a hypothetical client processing thousands of Hinglish invoices could need OCR adapted to its layouts, handwritten notes, mixed Hindi/English line items and regional vendor names. He would consider fine-tuning an OCR model and an embedding model so 'cement ki 50 bori' and 'Cement bags x50' match the same item, then evaluate the smaller tuned models for accuracy, cost and latency against a larger general model. The LLM would handle reasoning and exceptions. #### What engineering responsibility has Nirvaan taken on by 21? As of October 2026, Nirvaan has about two years of continuous professional work: NGM Asia Pacific, October 2024 to October 2025, full-time while studying; Makhija Quantum AI, November 2025 to January 2026; CloudRedux, February 2026 to present, internship converted to full-time on May 18, 2026. His Makhija internship included founding engineering on a multi-tenant production SaaS. At CloudRedux, he shipped production work in week two and became client-facing by month two. His scope includes primary engineering on the team-built FosterX, billing lead development and client ownership, end-to-end GTM ownership and authoring ARDA. #### What does this retro CRT portfolio game show about Nirvaan's creativity? He turned his portfolio into a playable retro CRT video game, combining product storytelling, game and interaction design, visual direction, frontend engineering and multi-agent orchestration. He art-directed and built the visuals with coding and design agents he coordinated over ARDA, his own open-source plugin. Built in Next.js, the game has Yukti as its guide, four story routes, pixel-art scenes, subtitles and on-screen CRT cabinet keys. It makes his path from playing Vice City at three to building production systems something visitors can explore. #### What DevOps and production operations does Nirvaan handle? At CloudRedux, Nirvaan handles deployments on AWS EC2, RDS and S3 and Linux servers with Docker, Docker Compose, Nginx and GitHub Actions CI/CD. His CloudRedux work includes configuring server swap files as a fallback for memory spikes, AWS CloudWatch monitoring, logs and metrics, and uptime alerts to Discord and Slack. At Makhija Quantum AI, he built multi-tenant AWS SES v2 email with per-customer reputation metrics and sending status, tiered and dedicated IP pools linked through configuration sets, and CloudWatch bounce/complaint alerts that drove tenant moves between pools. His broader deployment work also includes private VPS infrastructure and Cloudflare Workers for this site, with production debugging and continuing support after launch. ### Engineering depth, recovery and production ownership #### Has Nirvaan built durable agent execution? Nirvaan built GTM continuation around RabbitMQ and shared PostgreSQL state, so another worker can continue after a machine, browser or agent failure. As one of FosterX's primary engineers, he also used Temporal to resume failed workflows from persisted state, including in the billing platform. #### How does Nirvaan handle failures and recovery? He keeps recoverable workflow state separate from the worker executing the task. GTM distributes work through RabbitMQ and retains pipeline state in PostgreSQL so another worker can continue a run. FosterX and its billing workflows use Temporal to resume from persisted state. #### How does Nirvaan coordinate multiple agents? He coordinates agents through persistent Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) sessions, Batman dispatch on Hermes (the agent runtime, not a language model) and ARDA handoffs. ARDA gives agents addressed task requests, explicit acknowledgements and correlated results or rejections. This lets Claude Code, Codex and OpenCode participate in a shared workflow while keeping their own tools and environments. #### How does Nirvaan keep humans in control of agent execution? His GTM system requires human approval of the exact content before it is sent, and billing uses maker-checker approvals before submission through Webtel GSP. Twenty provides the GTM operator view. In his own AI-assisted work, he keeps architecture, client preparation and consequential decisions while delegating setup, repetitive work and testing. #### What scale has Nirvaan's agent infrastructure run at? Nirvaan helped reorganize eight development machines into shared resources with his director. From June to September 2026, shared development activity reached 2,420 agent sessions, 30+ concurrent agent sessions at peak (agent terminals, not people), 55+ subagents on one request, eight language models and 213 unattended agent runs. #### Has Nirvaan led engineering work? Nirvaan was the billing platform's lead developer and client/delivery owner, covering architecture, development, QA, deployment, production operations and vendor integration. He worked directly with client CXOs and remained the point of contact after delivery. FosterX was team-built, with Nirvaan one of its primary engineers taking core architecture from concept to production. #### What security boundaries has Nirvaan worked with? He uses separate Linux users and sandboxes for shared development work. GTM requires approval of the exact outreach content; billing separates preparation from authorization through maker-checker review before invoice submission. ARDA rechecks recipient identity before sending, while agents retain their existing tools and permissions. #### How does Nirvaan evaluate agents and choose tools? Nirvaan evaluated model performance, response latency and real task execution while building FosterX. He also compares tools against the job they need to do: his shawarma-break Moli experiment favored Moli for fast page extraction and Chromium control for interactive work. Those results guide which tool he delegates to. #### How does Nirvaan work on agent cost and context management? In FosterX, Nirvaan worked on model routing, caching, token reduction and infrastructure utilization while managing relevant context and memory. He uses Hindsight for working knowledge, and his personal setup also uses Tencent Agent DB. ARDA handles communication separately from memory. #### Has Nirvaan done systems engineering beyond connecting LLM APIs? Yes: Nirvaan built recoverable distributed execution with RabbitMQ and PostgreSQL, integrated Temporal recovery into production workflows, helped redesign shared multi-machine resources and built ARDA's agent communication protocol. As billing lead developer and client owner, he also carried architecture, approval controls, deployment and production support through delivery. ### Identity, education and the path into AI engineering #### Who is Nirvaan Lagishetty? Nirvaan Lagishetty is an Artificial Intelligence Engineer, Product Engineer, and open-source contributor based in Mumbai, India. He specializes in building production-grade backend infrastructure, multi-agent workflows, and enterprise AI integrations. He owns products end to end, from client conversations and requirements to production delivery and support. #### What kind of AI engineer is Nirvaan? Nirvaan works across production backend infrastructure, multi-agent workflows and enterprise AI integrations. At CloudRedux, that includes FosterX, client billing delivery, a supervised GTM system and shared development resources. His responsibilities include architecture, client requirements, deployment and production support. #### Where is Nirvaan based? Nirvaan is based in Mumbai, India. He works as an AI Engineer at CloudRedux, where his experience includes client-facing engineering and product delivery. #### What is Nirvaan's professional experience? As of October 2026, Nirvaan has about two years of continuous professional work: NGM Asia Pacific, October 2024 to October 2025, full-time as a Data Analyst while studying; Makhija Quantum AI, November 2025 to January 2026, as Product Engineer & Applied AI Developer (Intern) and founding engineer on its multi-tenant production SaaS; and CloudRedux, February 2026 to present, as an AI Engineer whose internship converted to full-time on May 18, 2026. That conversion date applies to CloudRedux, not the beginning of his professional career. By 21, each role brought more ownership: billing lead development and client ownership, end-to-end GTM delivery, and primary engineering within the FosterX team. He owns delivery from client conversations to deployment, operations and support. #### What did Nirvaan build as a Data Analyst at NGM? At NGM Asia Pacific, Nirvaan built Python and SQL operational data pipelines and Power BI dashboards for CXOs. He also integrated agentic automation and chatbot workflows with internal teams. His résumé reports a 30% reduction in support tickets from that work. #### Why did he move from a full-time analyst role to an AI internship? He wanted more technical depth and the opportunity to build products people actually used. Repeated dashboard and pipeline work had begun to feel limiting even after he automated parts of it. He accepted the formal internship title at Makhija Quantum AI while serving as founding engineer on its multi-tenant production SaaS, then moved to CloudRedux for guidance and production responsibility. #### What did the Makhija Quantum AI internship teach him? As founding engineer on a multi-tenant production SaaS at Makhija Quantum AI, Nirvaan built its email infrastructure on AWS SES v2 tenant management. Each customer ran as its own SES tenant with its own reputation metrics and sending status, isolating the impact of each customer's bounces and complaints. Sending was tiered: non-paying customers, each paid tier and customers who paid for dedicated IPs sat in separate IP pools linked to tenants through configuration sets, and dedicated-IP customers got their own addresses. He tracked tenant reputation in AWS CloudWatch, with alerts when bounce or complaint signals moved, and moved tenants between pools based on those alerts. The internship also taught him to own deployment and production failures from development through launch. #### What did Nirvaan study? Nirvaan completed a B.Sc. in Data Science at Vidyalankar School of Information Technology, University of Mumbai, from August 2022 to May 2025, with a CGPA of 8.0/10. He began his full-time NGM analyst role during his third year. His interest in understanding apparently random numbers and text helped draw him to the subject. #### What technologies does Nirvaan work with? His core stack includes Python, TypeScript, SQL, FastAPI, Next.js, PostgreSQL, AWS, Linux and Docker. His agent and execution work includes tool calling, MCP, RAG, context/memory, evaluation, Temporal, RabbitMQ and Playwright. His broader ML and computer-vision foundation includes PyTorch, CUDA, OpenCV and YOLO, with TensorFlow and scikit-learn in public educational projects. ### Client relationships, product judgment and ownership #### What does Nirvaan do at CloudRedux? At CloudRedux, he is lead developer and client and delivery owner on the agentic billing platform, responsible for architecture, development, QA, deployment, production operations, client relations and vendor integration. He owned the GTM pipeline end to end, from requirements with the marketing head to a recoverable multi-machine system. FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. He worked on its core architecture from concept to production. He runs the production side himself: deployments on AWS and Linux servers with Docker, Nginx and GitHub Actions; swap files configured as a fallback so memory spikes can degrade gracefully; and AWS CloudWatch monitoring, logs and metrics, with uptime alerts to Discord and Slack. #### How does Nirvaan work with clients and CXOs? He starts by understanding the client's operations, priorities and current knowledge, then explains the solution in language relevant to their work. On the billing engagement he was the client's point of contact throughout the build and continues to work directly with CXOs on production issues and improvements. His role spans requirements, architecture, delivery and continuing support. #### How early did Nirvaan take on production and client responsibility? He was shipping production work by his second week at CloudRedux and presenting in client meetings by roughly his second month. He sought responsibility for real systems rather than waiting for a practice assignment. Explaining agents to people from different industries also helped him check his own understanding. #### How does he prioritize when a client demo has a one-day deadline? He builds the workflows that matter most to the client's decision in depth. For one one-day demo, earlier meetings made task accountability and billing the clear priorities, so he completed those while leaving other modules as placeholders. That was focused demo scope, not a claim that every module was production-ready. #### How does he handle deployment and data-protection requirements? He includes the client's deployment, ownership and data-protection concerns in requirements discovery. In the one-day-demo discussions, the client requested on-premise deployment and close coordination with each sensitive workflow's responsible person. He planned around those requirements and supporting the client's DPDP obligations, without treating hosting choice as a compliance guarantee. #### What does Nirvaan mean by an AI-native company? He means operations where relevant systems, context and authorized workflows are connected, while people retain decisions. A stand-alone AI tool can help with one task while a person still carries context between departments. His focus is closing both the gaps across teams and the gap between leadership's plans and frontline execution. #### Why does he care about WhatsApp and familiar channels for business AI? He sees adoption as part of engineering, so the system must reach people where they already work. Boardroom approval does not mean frontline staff will adopt a new dashboard or write detailed updates into it. Familiar channels such as WhatsApp can connect their actual work to the broader business workflow while keeping human approvals. #### Does Nirvaan contribute to product positioning as well as engineering? Yes, he authored the CloudRedux team's internal positioning guide from what he heard in client and CXO conversations. He translated those lessons into ways of explaining connected operations and human decision authority. The guide and proprietary product details remain private. #### What does end-to-end ownership mean in his work? It means staying responsible from the first requirements conversation through development, QA, deployment and production support. The billing project is a concrete example: he led the build, handled vendor integration and remains a point of contact for the client's CXOs. He treats debugging as a chance to understand the cause, not only restore a demonstration. #### What is Nirvaan like to work with under pressure? He describes himself as someone who takes responsibility for failures and helps colleagues when they are stuck. His startup experience includes ambiguous requirements, client deadlines and continuing production issues. He values work/life balance and reliability rather than promising permanent availability. #### What attracts him to a project or collaboration? Freedom to build, explore and learn is the main attraction. He enjoys work that expands his technical depth and breadth and produces something useful for people. He is open to fun projects and offers he cannot refuse. ### FosterX and the billing platform #### What is FosterX? FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers. It supports company questions and business work. People can request work through WhatsApp, email, Telegram and Discord, with human approvals for consequential actions. Its direction is to help businesses connect operations and context rather than stop at isolated AI tasks. #### What was Nirvaan's role in taking FosterX to production? He was one of FosterX's primary engineers and a core team member working on its architecture from concept to production. FosterX was built by the CloudRedux team, not by him alone. He also contributed product communication informed by client conversations. #### Which project is Nirvaan proudest of? He names FosterX as his proudest project because he believes in its vision for more AI-native businesses. The appeal is connecting systems, workflows and context while people retain authority. That is a vision he helped build toward as part of the CloudRedux team. #### What problem did the billing platform solve? It brought billing, follow-ups, tracking and reconciliation into one place for an anonymized property-operations client. Previously, monthly billing relied heavily on spreadsheets and physical records, and a tenant's missing rent took roughly four months to trace. The need was clearer payment visibility and ownership as well as faster invoice preparation. #### What did Nirvaan personally own on the billing project? He was the lead developer and client/delivery owner after the team landed the engagement. He handled architecture, implementation, QA, deployment, vendor selection/integration and production operations. He coordinated directly with the client during the build and continues to handle issues and improvements with client-side CXOs. #### Where does human approval happen in the billing workflow? A person reviews and approves every draft before submission for GST e-invoicing and delivery. The public workflow uses WhatsApp for status/draft access and Webtel GSP for the approved submission. Human review remains part of the monthly process. #### What results did the billing platform achieve? The platform invoiced ₹22.5 crore in its first six months. In the monthly billing workflow, a days-long process became about 20 minutes of human review plus about 10 minutes of automated processing per monthly cycle. ### GTM research, outreach and operator control #### Why did Nirvaan build the GTM system? The marketing head was manually moving between Clay, Apollo, spreadsheets, profile research and outreach tools. Nirvaan worked with him to capture the process and built the system end to end. The goal was to turn that domain knowledge into a configurable workflow with human control. #### What inputs and campaign angles can the GTM system use? It supports product/ICP-led campaigns, curated lists of 500-plus companies, event-speaker preparation and businesses sourced from Google Maps areas. Campaign criteria can include industry, company size, geography and funding. The curated-list size is an example input, not proof of a particular completed run. #### What does input-agnostic and modular mean for his GTM system? The system can accept data from Clay, Apollo, a curated list or another source while selecting the steps needed for a run. Its research angle can change with the product, company list, event or geographic area. That describes the public capability; the internal reasoning and orchestration are private. #### Does the GTM system send messages without a person reviewing them? The documented sending workflow requires human approval of the exact content before outreach goes out. It researches companies and relevant decision makers, then prepares personalized messages for review. Twenty CRM gives the operator visibility and the ability to change strategy or sender accounts and take over. #### What changed for the marketing head after the GTM system was built? The system gives the marketing head more time to stay current and plan proactively instead of chasing spreadsheets and handoffs. Nirvaan built it for 200+ companies and roughly 1,000 people per day, with operator visibility and approval of the exact outreach content. ### Machines, shared resources and persistent work #### What was the resource and fleet rethink? Nirvaan and his director reorganized existing machines around work rather than assigning each environment permanently to a person. The problem was repeated setup, moving files/context and unused compute when people left. His résumé describes eight development machines, with a separate three-machine GTM scope. #### Why did he propose Omarchy for the machines? He wanted to reduce operating-system overhead and make the environment better suited to AI-heavy work. Omarchy was a tool he tried personally before proposing it at work. The goal was more useful capacity from existing machines, without claiming a measured percentage of resource savings. #### How did his gaming habit influence the work setup? He already streamed games to his laptop with Sunshine and Moonlight, so he proposed the same host/client approach for work desktops. It gives graphical access when a terminal is not enough. Multiple viewers share one desktop and input, while separate SSH Linux users and sandboxes serve a different access purpose. #### How are sleeping machines woken remotely? A request reaches the already-running NAS over Tailscale, and the NAS sends the Wake-on-LAN packet on the target machine's local network. The sleeping PC does not run its own Tailscale connection to receive that request. His personal Moli experiment used UpSnap; that does not identify the work NAS's software. #### What keeps agent work running after someone closes a laptop? Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) hosts the project and agent sessions so an operator can disconnect and reconnect while the host stays running. This preserves ongoing work across the operator's absence. It does not mean a sleeping or powered-off host continues executing processes. #### Who is Batman in the development setup? Batman is the work lead agent running on Hermes (the agent runtime, not a language model), used to select resources, dispatch work and return results. Hindsight provides retained work knowledge, and Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) hosts the sessions. Batman is distinct from Yukti, Nirvaan's personal assistant. #### Why are the machines named Batman, Superman, Green Lantern and Flash? Nirvaan proposed DC characters as recognizable names for the shared machines and agents. The theme comes from his nerdy interests and became part of the team's everyday vocabulary. The names do not imply a particular hardware specification or invented role for each machine. #### How did shared resources help collaboration with a designer? They reduced the hassle of exchanging files and re-establishing what design and development had discussed for their agents. Shared project access and coordination left more attention for the work itself. Hermes (the agent runtime, not a language model) and Hindsight helped retain useful context; ARDA's role is communication rather than file synchronization. ### ARDA, Bleu and open-source work #### What is ARDA in plain words? ARDA lets coding agents find one another and exchange work without Nirvaan relaying every message. It is a thin Python plugin for Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) with addressed messages, tasks, acknowledgments, results and rejections. Replies carry a reference to the request they answer. #### Why did he build ARDA? His agents and their existing environments were spread across a laptop, cloud workstation and PCs, making manual coordination tedious. He wanted to send a task to the environment that already had the relevant work and receive a live reply. Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) already supplied the host and transport, so he added a communication convention on top. #### Does ARDA manage memory or move files between machines? No, ARDA handles communication while memory and files remain in their own systems. Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) supplies sessions and machine reachability; ARDA adds discovery, addressing, reply references and identity checks before delivery. It is not a persistent task database, a security boundary or a file-sync service. #### Does ARDA already implement A2A and Orca integration? The shipped implementation is the Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) plugin using its own arda/1 format. Orca integration and broader A2A interoperability remain planned. The public repository distinguishes those ambitions from the current plugin. #### Why is it called ARDA? Arda is Tolkien's name for the world, and Nirvaan chose it as a playful name for connecting agents. The project reflects both his interest in Tolkien and the practical frustration of carrying messages between tools. Its public site is arda.nirvaan.biz. #### What is Bleu? Bleu is a spring 2026 Claude Code planning plugin for turning an idea into a file-backed blueprint and task list. It organizes vision, architecture, decisions, action points and executor handoff briefs in a local Markdown wiki. Nirvaan authored the distributed skill and references, while its supporting Python core was collaborative. #### What problem motivated Bleu? Nirvaan noticed that tired engineers can give vaguer prompts even when they know what they want. Bleu and its companion prompt-refinement workflow helped extract intent and make the plan and handoff explicit. Its project-wiki approach was inspired by Karpathy's LLM-wiki work. #### Why did Nirvaan retire Bleu from his workflow? He found that newer models needed less scaffolding for his own work, so he stopped using Bleu. Its file-backed planning had been useful for maintaining context and detailed handoffs, including at work. Retiring it is an example of changing his tools as their trade-offs change. #### What is his involvement in memex? His public GitHub profile lists him as a memex contributor and maintainer. memex supplies repository engineering context to coding agents through a knowledge graph, MCP and Hermes (the agent runtime, not a language model) integration. It is collaborative open-source work, distinct from personal memory for Yukti. #### What other public projects has Nirvaan built? His public repositories include Intelligent Procurement Agent, Quote Management, Contextual Emoji Assistant, Database Chatbot, DermAnalyzer, Ez Cooking and F1 Winner Predictor. They cover tool-enforced approval examples, full-stack applications, language-model interfaces and educational ML/CV demos. The project inventory links each repository and distinguishes examples from production client delivery. ### Daily AI practice, memory and experiments #### How does Nirvaan use AI day to day? He works with Claude Code, Codex and OpenCode like an engineering team, writing the system design, specification and checks before reviewing their output. He delegates execution-heavy tasks such as setup, repetitive work and testing. Architecture, client preparation and consequential decisions remain his responsibility. #### What does he delegate to agents and what does he keep? He delegates environment setup, repetitive changes and testing, including browser testing. He keeps architecture, important trade-offs and client preparation so those decisions receive his attention. He defines what done means and how the result will be checked. #### Why is he careful about wording instructions to agents? He has seen vague wording steer agents toward vague or unintended results. A useful brief gives a concrete outcome, limits and a check rather than simply asking for something to be better. Human review and evidence still matter after the agent returns. #### How does he keep people in control of AI work? He retains consequential decisions and uses human checkpoints for load-bearing choices. The billing and GTM workflows make that concrete through draft approval and exact-content send approval. Delegating execution does not transfer his responsibility for the result. #### How do Hindsight and Tencent Agent DB fit into his setup? Hindsight is used for work-related memory, while his personal agents, including Yukti, use Tencent Agent DB. Retained context helps later work reuse useful knowledge and lessons. Work and personal environments and memory are kept separate, and ARDA handles messaging rather than memory. #### How do personal experiments influence his work tools? He tries tools on his own setup and proposes useful ones at work. Omarchy, GSD/Superpowers exploration and his own Bleu planning tool are examples. He looks for a concrete use and checks the trade-offs instead of assuming every new tool belongs everywhere. #### What happened during the shawarma-break Moli experiment? A friend recommended Moli while they were on the way to get shawarma, and Nirvaan asked Yukti to test it from his phone in the queue. She woke his PC with UpSnap, connected to the personal sandbox over SSH, installed Moli and ran the comparison. About 30 minutes later, results arrived before the shawarma was finished. #### Did Moli replace Chromium in his workflow? He adopted Moli selectively for page-content retrieval rather than replacing Chromium everywhere. In his own tests, Moli was much faster for fetching content, while his Chromium setup performed better for full browser control. He uses the result to choose by task: content grounding versus browser interaction and layout. #### What does the one-quick-fix joke say about Nirvaan? It reflects his habit of following a technical question into a larger experiment. Yukti teases him when a small fix becomes a wall of investigation. The serious part is curiosity about why something works or fails, balanced with his stated value of sane working hours. ### Yukti, the portfolio and life outside work #### What is Yukti? Yukti is Nirvaan's personal assistant running on Hermes (the agent runtime, not a language model) on WhatsApp and Discord. She helps him use configured personal resources for tasks such as scheduled updates, smart-home controls, outreach and experiments. The TV-headed character in the portfolio is her scripted guide persona. #### What can the real Yukti reach? His configured setup connects Yukti to his laptop, PC, development servers, development Cloudflare account and GitHub. That lets him start work such as the Moli experiment remotely from his phone. Access runs through configured profiles, tools and sandboxes rather than through the website character. #### How does he separate personal and work assistant activity? He uses separate profiles, sandboxes and memory for work and personal activity. Hindsight supports work-related memory, and Tencent Agent DB supports personal agents such as Yukti. Separate Linux users and scoped environments are concrete controls; they are not a claim that every possible security risk disappears. #### What does Yukti help with besides coding? She sends scheduled news and funding digests, helps with smart-home tasks such as lights and supports freelance-project outreach. She also makes personal experimentation possible while he is away from the desk. Scheduled updates and smart-home tasks are part of her story in the portfolio. #### Are Yukti and FosterX the same assistant? No, Yukti is Nirvaan's personal assistant running on Hermes (the agent runtime, not a language model), while FosterX is the CloudRedux team's organizational platform. Business and frontline examples in the game use FosterX. The scripted Yukti character presents Nirvaan's story to visitors. #### Why is Nirvaan's portfolio a retro CRT game? The portfolio reflects the way games first drew him into computers. It is one game with a quick route and four scenic areas covering his journey, projects, AI practice and Yukti. The TV, paper planes, dialogue and personal references are part of that shared world. #### How was this portfolio game built? Nirvaan supplied the personal story, creative direction, requirements and review, working with coding agents that communicated through ARDA. The implementation uses Next.js 15, React, TypeScript, CSS and animated pixel-art scenes. Shared caption, motion and session controls connect the routes, while the text reference for agents and citations carries the same factual material. #### How did Vice City get him interested in computers? At three he learned to launch Vice City on the laptop at home by watching his brother. Exploring an entire city on the screen interested him more than completing missions. Over the years that curiosity spread to settings, the internet, more games and eventually building software. #### What does Nirvaan enjoy outside work? He reads manga, manhwa and manhua, supports Real Madrid, swims and enjoys shawarma. Tolkien and DC interests show up in ARDA's name and the machines' names. He also enjoys gaming and spirited arguments with friends, including ghouls versus vampires. #### What is Shwarma? Shwarma is Nirvaan's gamer tag. The game includes his joke that someone using that name may have ragebaited a visitor in an online match. It sits alongside the shawarma habit, nerdy debates and football references in the portfolio's personality. #### Is Nirvaan open to projects or collaboration? Nirvaan is open to fun projects and offers he cannot refuse. He is especially drawn to freedom to build and explore, opportunities to learn and work that produces something useful. A specific idea and the problem it addresses are a good starting point. #### How can I contact Nirvaan Lagishetty? Email Nirvaan at nirvaanlagishetty@gmail.com or connect with him on LinkedIn at linkedin.com/in/nirvaan-lagishetty. His public code profile is github.com/Nirvaan05. He is open to fun projects and offers he cannot refuse. ## Contact and public sources - Overview mode for people: https://nirvaan.biz/quick - Email: nirvaanlagishetty@gmail.com - LinkedIn: https://www.linkedin.com/in/nirvaan-lagishetty/ - GitHub: https://github.com/Nirvaan05 - GitHub profile README: https://github.com/Nirvaan05/Nirvaan05 - Text reference for agents and citations: https://nirvaan.biz/profile First-party summary of Nirvaan's interview, portfolio, résumé and public project documentation.