Full-stack engineer building
production AI-powered systems.

Production-grade FastAPI backends with JWT auth, rate limiting, Pydantic validation, and full Swagger docs. Built to handle real traffic from day one.

End-to-end webhook pipelines with HMAC signature verification, real-time event inspection via WebSocket, retry logic, and full audit trail.

Hybrid search combining dense vector embeddings with BM25 keyword matching and RRF re-ranking. SSE streaming results. Qdrant as the vector store.

Complete product from zero: FastAPI backend, Next.js frontend, PostgreSQL, auth, deployment on Railway + Vercel. GSAP / Framer Motion animations included.

Token-by-token streaming from OpenAI or Anthropic SDKs, wired through SSE or the Vercel AI SDK's useChat. Disconnect-aware: client drops the tab, the upstream call stops too.

Anthropic tool-use loops: model requests a function, the server executes it against a real backend, the result loops back until a final answer. Decoupled so swapping one tool touches zero loop logic.

Four-bucket cost accounting — input, output, cache-write, cache-read priced separately — checked before a request is even allowed to fire. Context-window auto-summarisation at a real, measured 80% threshold, not a guess.

Closing the gap between 'asked the model for JSON' and 'got JSON back.' Strict schema mode plus runtime validation at every checkpoint — the model's output is trusted exactly as much as a user's form input.

Swap GPT-4 for Claude via one environment variable, not a rewrite. Paired with retry/backoff logic that's tested to actually fire correctly — not retry code that looks right and silently does nothing.