AI Engineer — agentic systems, multi-agent research, and the full stack around them. Two products I built solo have 11,000+ organic users. Code merged into a 15k-star open-source agent platform. One published benchmark on how agents resolve memory conflicts.
B.S. (Honors) in Data Science & AI, IIT Guwahati.
LinkedIn · Portfolio · LeetCode · Email
Research Intern — RAIVN Lab, University of Washington (Remote) Multi-agent web browsing: up to 4 vision-language agents drive isolated Playwright browsers in parallel, best-of-N over a pluggable judge. Swappable policy layer across MolmoWeb-4B, Qwen-VL and Qwen3.5-27B — a 6-target grounding gate took the live adapter from 0/6 to 6/6 clicks by catching a coordinate-units bug before it read as model error.
Agentic AI Engineer — Rowboat Labs (YC S24) (SF, Remote) Shipped Skills (agents load instruction sets and tools on demand, as installable packages) and ChatGPT OAuth 2.0 + PKCE into a 15k-star open-source agent platform. 30+ PRs merged, both co-founders reviewing.
ThinkSpace AI (Singapore) — Founding engineer on a multi-document research assistant in Electron: sentence-window RAG with page-coordinate-anchored citations across PDF/DOCX, Drive and live web, validated by a legal QA team on a ~100-question gold standard.
Compeers AI (US) — Two market-research tools for a B2B intelligence platform: a 4-module SWOT pipeline (Search → PDF/CSV → SEC EDGAR → Trends) and a Reddit audience profiler.
CareerLift — 10,000+ organic users LLM agents parse your resume, semantically match it against live job listings and 1,500+ IIT professor research roles, and return ranked roles with skill-gap analysis. A jobs pipeline refreshes 3,500+ openings every 12–24 hours with no manual curation.
Drona AI — 1,000+ organic users An interview agent that generates role-specific questions from an uploaded PDF, tunes difficulty on a rolling performance window, and streams back a feedback report. Re-architected from Streamlit to serverless Next.js, now running at zero infra cost.
OpenCollab MCP — open source, MIT
An MCP server for open-source contributors: skill-matched good first issues, repo health scoring, PR plans. Zero-infra deploy via uvx.
Navigating Epistemic Parity in LLM Agents — sole author, Zenodo preprint, July 2026. A 52-scenario benchmark on what happens when an agent's injected memory contradicts its SKILL.md file. Across 1,383 graded runs there was no default precedence, and the agent resolved the conflict silently in 98.8% of tool-call runs.
Python, TypeScript, SQL · LangGraph, LangChain, MCP, RAG, FAISS/ChromaDB, PyTorch, Hugging Face, vision-language models, LLM evals · FastAPI, Node.js, PostgreSQL, Supabase, Redis, Docker · React, Next.js, Tailwind, Electron, Vercel
LeetCode 1710 peak (top 16.6%), 1000+ problems solved.
Demos are easy; products people come back to are hard. I optimise for the second one.
Open to AI engineering roles. Remote preferred, flexible on location.