AI Engineer | RAG Pipelines, LLM Integration & Production-Ready Generative AI
I build AI features that hold up in production — document Q&A, semantic search over your own data, image and video generation, streaming chat, and the async pipelines that keep all of it from falling over.
Most developers can call an API. The hard part is what happens after: a video generation that takes 90 seconds, a webhook that fires twice, a RAG retrieval query that returns confidently wrong chunks, a model bill that triples overnight. That's the layer I build.
WHAT I BUILD
- Document Q&A & Semantic Search — PostgreSQL + pgvector, hybrid retrieval, chunking tuned to your document type, citations on every answer. No vector database subscription, no framework lock-in.
- Generative AI Features — video, image, and audio via Replicate and other model APIs, wired into your existing product
- Async Job Pipelines — queuing, retries, idempotent webhook handling, checking progress, graceful failure
- Streaming AI Chat — SSE with proper cancellation, token handling, and reconnection, inside your Next.js or React app
- LLM Cost Control — usage metering, per-user limits, model routing, caching, so your inference spend stays predictable
- The Full Application — Next.js frontends, Node.js and FastAPI backends, PostgreSQL with Prisma, AWS deployment
RECENT WORK
AIVideosTube — an AI video generation platform serving 18,000+ users, built end to end. Users submit a prompt, a job runs across Replicate and other model APIs for a minute or more, and the result lands reliably in their library. Behind it: a queue that survives failed inference calls, webhook handling that stays correct when providers deliver twice, per-user usage limits, and cost controls that hold unit economics steady as usage grows.
Pakistan Law Bot — an AI legal research platform. I built the full application around the retrieval layer: authentication, subscriptions, the API layer, and a streaming answer interface delivering responses token by token with source citations attached.
HOW I WORK
On retrieval projects, I test against your real documents and your real questions before writing any application code. RAG fails on retrieval quality far more often than on the model — if the wrong chunks come back, no prompt engineering saves the answer.
On generation work, I ask what happens when the model call fails before I ask which model you want. AI features look finished in a demo and break in production — timeouts, duplicate charges, jobs stuck in "processing" forever. I build for that from the first commit.
Six years full-stack, so the AI feature ships inside a working product — not as a prototype someone else has to productionize.
Tell me what you're building and what your users will actually do with it, and I'll tell you what it takes to run it in production.