• 🔭 Currently an AI Engineer on the Agent Platform team at Dwarf Star Systems, Bengaluru.
• 🤖 I build production RAG assistants that don't just answer questions — they take actions across platform modules through tools.
• 🎓 Dual Degree (B.Tech + M.Tech), Electrical Engineering — IIT Bombay.
• 🛠 Experienced across the whole path: LangGraph agents, FastAPI backends, pgvector retrieval, MLOps and Kubernetes rollouts.
• 📏 I care about evals over vibes — retrieval recall measured apart from answer quality, trajectory-level regression suites gating every change.
• 🔐 Big on human-in-the-loop confirmation, row-level security and audit trails — agents that touch real data should ask first.
• ⚡ Happiest when a messy system becomes a measurable, observable, reliable one.

• AI Engineer, Agent Platform @ Dwarf Star Systems, Bengaluru — Jan '26 – Present
• MLOps Engineer, Model Delivery Platform @ QualityKiosk Technologies, Mumbai — Jul '25 – Jan '26
• Applied AI Engineer, Enterprise Assistant @ Kotak Mahindra Bank, Mumbai — May '24 – Jul '24
• NLP & Retrieval Engineer, Biomedical Search @ BenchSci, Toronto — Jun '23 – Aug '23

I build agentic RAG systems — assistants grounded in real data that can reason, retrieve and act — and the evaluation and delivery infrastructure that keeps them honest in production.

• Agentic RAG: LangGraph agent layers over FastAPI, SQLAlchemy and pgvector, unifying data fragmented across a platform, served over gRPC with SSE streaming.
• Grounded, refusing answers: citations on every claim and forced refusal on thin context — cut unsupported answers by 80%.
• Evaluation harnesses: scoring retrieval recall separately from answer quality (71% → 93% recall), plus trajectory-level regression suites that gate every deploy.
• Safety & tenancy: human-in-the-loop confirmation on destructive actions, row-level security for tenant isolation, immutable audit trails.
• MLOps: automated retraining on Airflow + MLflow, PSI drift detection with Grafana dashboards, Kubernetes canary rollouts with automatic rollback on p99 breach.
• Model efficiency: matched a hosted API baseline on accuracy at 1/12th the cost via an 8B QLoRA fine-tune served on vLLM.

Task Tracker — Full Stack · Live App

A multiuser task tracker with an AI assistant built in. FastAPI + SQLAlchemy + Alembic on Neon Postgres, React (Vite) frontend, deployed on Render. GitHub Actions CI/CD runs migrations and the full test suite against Postgres on every push, and production deploys are gated on that suite.

Copilot — Agentic RAG

The LangGraph conversational agent powering the task tracker's assistant. RAG over user tasks with pgvector and Gemini embeddings, human-in-the-loop confirmation for destructive actions, per-user isolation, rate limiting, audit logging, and an automated eval suite that regression tests the agent on every change.

Languages, Frameworks & Data

AI & Retrieval

DevOps & Observability