Work
First engineer at a startup, platform engineer at scale, LLM engineer on frontier models, and now a founder. Tap any company to see what I built there, or read the full history on LinkedIn.
-
2026–nowAxivion LabsIndependent AI Product Engineer (Founder)Production AI systems for startups: a multi-agent deal desk, an enterprise RAG assistant that took eval accuracy from 61% to 89%, and an agentic content pipeline. Details →
Jan 2026 – Present · Bengaluru
Building AI products and services. I ship our own AI products and build AI systems end to end, including problem discovery, architecture, deployment, and evaluation.
What I did
- Deal desk: an AI sales team for inbound leads. Built a multi-agent system that qualifies leads, researches prospects, recommends pricing, and drafts personalized replies. It handles first-touch qualification, follow-ups, and pricing negotiation across 1500+ leads a month, saving ~25 hours a week of manual sales-ops work.
- Knowledge assistant: answers you can trust. Shipped an enterprise assistant over 12K documents with hybrid retrieval, re-ranking, metadata filters, and role-based access. Answer accuracy rose from 61% to 89% on a 200-question eval set, with Langfuse tracing and cited sources.
- Content pipeline: from keyword to published post. Designed an agentic pipeline for a MarTech client covering keyword research, planning, writing, SEO, CMS publishing, internal linking, and distribution. Output grew from 8 to 60 pieces a month at ~30% of the previous cost per piece.
- Evals and plumbing: the unglamorous half. Built the agent evaluation pipeline on Langfuse that traces every run and regression-tests task success, tool-call accuracy, retrieval relevance, and answer grounding before a prompt change ships. Underneath, FastAPI services on Amazon Bedrock handle inference, async jobs, and tool execution with rate limits, retries, and structured outputs.
Stack
PythonFastAPILangGraphClaude / OpenAIMCPRAGpgvectorPostgreSQLRedisLangfuseAmazon BedrockDocker
-
2024–2025WrikeSenior Software Engineer II, AILed AI Portal, an internal RAG system serving 10k+ queries a day, and built Herald, an MCP server that feeds company context to AI coding tools. Details →
Nov 2024 – Dec 2025 · Bengaluru
Wrike makes work-management software. I built the AI tools its own teams used every day, and shipped full-stack features on the core product.
What I did
- AI Portal: ask the whole company a question. Led the build of an internal RAG system on Google Vertex AI that answers questions across company knowledge, codebases, and support docs. It served 10k+ queries a day and cut the time it took people to find answers by 65%.
- Herald: company context for AI coding tools. Designed and shipped an MCP server that gives AI-powered IDEs our project guidelines, review standards, internal docs, and coding conventions, so their suggestions follow the team’s own standards. Through the ESLint 9 upgrade, the new dialog system, and the IconV2 migration, it cut the frontend team’s manual work by 90%.
- App Marketplace: safe doors for third-party apps. Built the integration services behind Wrike’s App Marketplace: OAuth 2.0 authorization, webhook-based event processing, and REST connectivity, so outside apps could plug into customer workspaces securely. Also shipped core-product features end to end and mentored developers on engineering practices.
Stack
Vertex AIRAGMCPOAuth 2.0TypeScriptReactNode.js
-
2024–2025Turing ContractLLM EngineerCoding evaluations, code review, and failure analysis for Meta’s Llama 4 and OpenAI’s GPT models with frontier-model research teams. Details →
Sep 2024 – Jun 2025 · Contract
Worked with frontier-model research teams through Turing on Meta’s Llama 4 and OpenAI’s GPT models, focusing on improving their ability to generate, debug, and review code. My work combined coding evaluations, technical code review, and failure analysis to identify where models struggled and help improve their responses.
What I did
- Designed coding tasks and evaluations covering code generation, debugging, code review, and completion. Developed and refined 500+ prompts to test model reasoning and the correctness of generated solutions.
- Reviewed model-generated code for correctness, logical errors, edge cases, and adherence to requirements. Provided detailed technical feedback explaining what failed, why it failed, and how the solution should change.
- Ran systematic model evaluations using human assessment and automated metrics, including pass@k, to compare responses and identify recurring coding failures.
- Developed and refined 1,000+ coding examples, working with AI researchers and engineers to incorporate evaluation findings into coding-focused post-training and RLHF workflows.
Models & methods
Meta Llama 4OpenAI GPTGeminiLLM evaluationCode reviewPrompt engineeringRLHFpass@k
-
2021–2024AirmeetSenior Software EngineerCore platform for live events at 100k+ concurrent attendees. Led the GenAI initiative: event generation, session summaries, and real-time translation. Details →
May 2021 – Oct 2024 · Bengaluru
Airmeet is a platform for virtual and live events. I was on the core platform team, working on architecture and performance, and led the company’s GenAI initiative.
What I did
- Core event platform: built for big crowds. Owned the architecture of the microservices behind every event, sustaining 100k+ concurrent attendees at sub-100ms session latency on AWS Lambda, Redis, and Kafka.
- GenAI initiative: from proposal to production. Led the rollout of GPT-4 event generation (60% faster event setup), automatic session summaries (500+ hours a month), Whisper-based live translation and transcription, and semantic matching for networking rooms.
- Performance program: faster events for everyone. Cut event load times by 45% with code splitting and tree shaking, improved FCP and LCP across the product, and set the frontend performance standards the team built on.
- On the line during live events. Diagnosed and fixed production issues across Firebase, backend APIs, and the real-time systems while events were running, often working directly with enterprise customers so their event went out without a hitch.
Stack
ReactTypeScriptNode.jsJavaReduxKafkaRedisAWS LambdaFirebaseGPT-4Whisper
-
2020–202191WheelsFounding Software EngineerFounding engineer working with the CEO: built the marketplace end to end (frontend, backend, and AWS), an in-house CRM, and lead integrations, then moved the site to Next.js with Elasticsearch search. Details →
Jan 2020 – May 2021 · Gurugram
91Wheels is an automobile marketplace for listings, reviews, and leads. I joined as a founding engineer, worked directly with the CEO on what to build, and built the platform and its internal tools from the ground up.
What I did
- Founding engineer, working with the CEO. Worked side by side with the CEO to turn business goals into product: deciding what to build first, scoping features, and shipping them end to end.
- The whole stack, from scratch. Built the listing, review, and lead-generation platform end to end: the frontend in React, the backend and REST APIs in NestJS with MySQL and Redis, and the entire AWS setup it ran on.
- An in-house CRM. Built the CRM and lead-management modules that hold every vehicle and lead record, and that serve as the backend for the frontend.
- Lead integrations. Integrated lead sources into the platform so incoming leads flowed straight into the CRM, without manual entry.
- Faster, findable pages. Rebuilt the site on Next.js for server-side rendering, and added Elasticsearch search across every vehicle and page.
Stack
ReactNext.jsNestJSExpressMySQLRedisElasticsearchAWSWordPress
-
2019–2020FluxAuto (YC 2018)Software Engineer (Part-time)Part-time engineer: built the live GPS dashboard for the test fleet, and the internal tools and APIs that automated image annotation for the perception team. Details →
Nov 2019 – Sep 2020 · Bengaluru
FluxAuto is a Y Combinator 2018 startup in Bengaluru building self-driving technology for trucks. I worked there part time as a software engineer, building the web dashboards and backend services the engineering team used every day: tracking vehicles on the road, and preparing the image data their models trained on.
What I did
- Live vehicle tracking. Built a React dashboard that plotted each vehicle's GPS position on a map in real time, so the team could follow test runs live and review where a vehicle had been.
- Tracking backend. Wrote the Node.js and NestJS services that took in the GPS stream from the vehicles, stored it in SQL, and served it to the dashboard through APIs.
- Annotation automation. Built internal dashboards and API endpoints for the image-annotation workflow: getting camera images to annotators, tracking progress, and collecting the labels, replacing steps the team used to do by hand.
Stack
ReactNode.jsNestJSTypeScriptAWSSQL
Before all that: B.E. in Computer Science, Galgotias University, 2015–2019.