Building AI thatshipsto production

The core of most of what I build. I work with RAG and agentic systems on LangGraph and LangChain, pairing GPT-4o and Azure OpenAI with ReAct agents, careful prompting, and local models through Ollama when it makes sense.

Good answers come from good retrieval, not a clever prompt. I reach for pgvector day to day, with FAISS and ChromaDB where they fit, and spend the real time on chunking, semantic search, and killing near duplicate results before they ever reach the model.

Once something is live you have to see what it actually did. I trace every run with Langfuse and Arize Phoenix, watch prompt and token behaviour in production, and build in defenses so a public facing bot holds up against prompt injection.

Most of my systems are Python and FastAPI, with Celery handling the work that should not block a request. I also build in NestJS and Node when a project calls for it, across REST APIs and streaming responses over SSE.

I build the front ends too, mostly Next.js and React in TypeScript with Tailwind. I care about the small things that make an interface feel right, from smooth streaming to state that stays out of your way, and I have shipped mobile in Flutter.

The unglamorous layer everything sits on. PostgreSQL is my default, with Redis, MongoDB, and MySQL where they earn their place, Prisma over the schema, and Docker plus AWS S3 to ship it and keep it running.