Self-taught Full-Stack Agentic AI Engineer who builds production LLM applications end to end — React/Next.js/TypeScript frontends, Python (FastAPI) and Node.js backends, Claude & OpenAI APIs — with the rigor layer most portfolios skip: retrieval grounding, measured evaluations, tracing, and infrastructure as code. My portfolio site is itself a live production system that runs every capability it claims. Seven years of client-facing technical delivery leadership mean I’m equally at home building the product and demoing it to customers.
A site that runs its own claims instead of listing them — eight shipped modules, each documented and defensible. RAG assistant with citations (pgvector) that answers only from real documents and refuses what they don’t cover; a 108-case golden-dataset eval suite scoring 0.992 retrieval recall@5 and 92/92 valid citations at $0.41 per run, which caught a live overclaim bug that was fixed, re-measured, and redeployed; request tracing with a public /metrics endpoint; a tool-using agent over portfolio data; an open-source MCP server; token streaming; Docker; and Terraform that round-trips clean against live infrastructure. Typed FastAPI backend (Pydantic, pytest, OpenAPI), CI/CD rebuilding two clouds on one push.
Custom support assistant on the Claude API, grounded in a product’s real knowledge: answers only from known content, refuses honestly instead of hallucinating, and hands off cleanly to a human. React + Vercel serverless, API key held server-side.
Framework-free responsive landing page built from scratch: CSS design tokens, fluid type with clamp(), auto-fit grid. Deployed with push-to-deploy on Vercel.