I build AI systems that survive
contact with production.

An OpenAI-compatible gateway that sits between your services and model providers.

Speaks the OpenAI wire format and rewrites the model field before forwarding, so callers stay backend-agnostic. A middleware pipeline handles per-IP rate limiting, structured request logs, and atomic cost accounting, with primary → fallback routing when a provider returns 5xx or times out. Standard library only — no external dependencies, one binary.

• Go
• stdlib only
• LLM infrastructure
• Failover

End-to-end retrieval-augmented generation on Google Cloud.

Documents and queries are embedded asymmetrically — RETRIEVAL_DOCUMENT for the corpus, RETRIEVAL_QUERY for the question — which is what actually moves precision with text-embedding-004. Vectors live in BigQuery behind an IVF index that approximates at scale and falls back to exact search on small corpora, so recall does not quietly degrade in development.

• Python
• Vertex AI
• BigQuery Vector Search
• RAG

An LLM-as-judge evaluation framework built to gate a deployment pipeline.

Faithfulness and correctness scorers extend a common LLMJudge base, so a new metric is a subclass and a prompt. A batch runner routes the right keys to each evaluator and exports CSV/JSON — then exits non-zero the moment any pass rate drops below its threshold, which is the part that makes it usable in CI rather than as a notebook.

• Python
• Gemini
• Evaluation
• CI gating

A ratings and reviews service in the same problem domain I now work in professionally.

Ports-and-adapters layout with two deliberately separate HTTP surfaces — a public one and a moderation one — so no route ever crosses between them by accident. Reviews move through an explicit PENDING → APPROVED | REJECTED state machine, idempotency keys absorb retry storms, and the rating aggregate is recomputed on approval behind a seam designed to become an async consumer.

• Go
• Hexagonal architecture
• REST
• Idempotency