Full-Stack AI · Backend · GenAI Engineer
_{Bangalore, India · open to relocation / remote}

Full-stack AI engineer who owns systems end to end — backend-first.

I have shipped two live production AI systems solo: a 16-engine AI platform with hybrid RAG, a LangGraph multi-agent research assistant and LLMOps instrumentation; and a conversational analytics platform built on a deterministic engine/AI separation. Python/Django foundation, backed by formal GenAI engineering training (IBM Professional Certificate). I design around engine isolation, async workers and vector retrieval.

I spent 2019–2024 preparing full-time for India's Civil Services Examination, then GATE through early 2025. I returned to software engineering in 2025 — and the first thing I built was the platform I wish I had while preparing.

Live: theknowledgeorbits.com · Built solo, end to end

• Architected a modular monolith of 16 isolated Django domain engines — each owns its models and communicates only via internal APIs, giving independent domain ownership without the operational overhead of microservices.
• Built a hybrid RAG pipeline — semantic (pgvector cosine) and lexical (BM25) retrieval fused via Reciprocal Rank Fusion, with a relevance gate and TopicRelation graph expansion.
• Built the research_agent — a 7-node LangGraph multi-agent workflow running in a django-background-tasks worker to bypass Render's 30s HTTP timeout, streaming live progress to the frontend over SSE.
• Integrated LLMOps and AgentOps: Langfuse tracing, DeepEval evaluations, a Redis-backed rate limiter, structlog logging, Sentry tracking, and LLM pooling across multiple providers.

Built solo, end to end

• Designed the system so a deterministic Pandas/SQL engine handles all computation, with the LLM used only to interpret intent and narrate results — keeping analytical output reproducible.
• Built a multi-key LLM pool (round-robin with per-key cooldown) and a 5-layer input-validation guard behind an SSE streaming chat interface.
• Solved real production failures: DB connection exhaustion on serverless, OOM on 51k-row datasets, and a WIF auth failure caused by repo-name case sensitivity.
• Deployed on GCP Cloud Run with keyless CI/CD via GitHub Actions and Workload Identity Federation.

Built solo, end to end

• A Planner routes each question to a cheap single-search path or fans out parallel Researchers; a Synthesizer merges cited findings or abstains; a Critic verifies every claim against retrieved evidence before a human approves the answer.
• Dynamic parallel fan-out via LangGraph's Send API with bounded retry and escalation cycles — every loop hard-capped in code — plus human-in-the-loop approval with durable on-disk checkpointing that resumes exactly where it paused.
• Citations validated in code against the chunks actually retrieved; fabricated or out-of-context citations are rejected, forcing an honest abstain. Quality scored with RAGAS.

Languages & Core   Python TypeScript SQL

Backend   Django DRF FastAPI Flask REST APIs RBAC Async workers

Frontend   Next.js React Tailwind CSS shadcn/ui React Flow

AI / GenAI   LangGraph Hybrid RAG Embeddings Prompt Engineering Langfuse DeepEval RAGAS Tool Calling

Data   PostgreSQL pgvector FAISS Redis Pandas NumPy Scikit-learn

Infra & DevOps   Docker GCP GitHub Actions Render Vercel Supabase Sentry structlog Linux

• IBM Generative AI Engineering Professional Certificate — Coursera, 16 courses
• IBM RAG and Agentic AI Professional Certificate — Coursera, 10 courses (in progress)
• MLOps Bootcamp (10 end-to-end projects) · GenAI with LangChain & HuggingFace · Data Science / ML / DL / NLP Bootcamp — Udemy
• Python for Everybody Specialization — University of Michigan · Python & Statistics for Financial Analysis — HKUST
• 🏅 Qualified GATE 2025 — Computer Science & Engineering
• 🏅 Qualified GATE 2025 — Data Science & Artificial Intelligence
• 🏅 AIR 4048 — GATE 2017, Computer Science & Engineering

B.Tech, Computer Science & Engineering — Jaipur Engineering College and Research Centre

Everything above is built and maintained solo, in the open. If any of it helped you, you can buy me a coffee — entirely voluntary.

_{Backend-first. Production from day one.}