Lghoumaigui Y.

I build and fix production AI integrations — Claude and OpenAI APIs running on AWS, where reliability and cost matter more than demos. Day to day I work on production services across most of the AWS catalogue, with AI tooling used heavily throughout. Alongside that I build and run my own products, so I write integration code as well as operate it. Most of my work is the unglamorous half. The integration already exists, it mostly works, and it is slow, expensive, or failing in a way nobody has traced yet. I find out why, with logs and evidence rather than guesses. What I do: • LLM integration — Claude API, prompt handling, structured output, streaming, retries and backoff • Cost work — prompt caching, batching, model and effort selection, token profiling against real traffic • AWS — Lambda, API Gateway, IAM, CloudWatch, and most of the rest of the catalogue • Guardrails — structured logging with traceable request ids, tests at the integration boundary, eval sets so prompt changes are measured, alerting on error rate, latency and spend • Backend — TypeScript/Node, Python, PostgreSQL, REST APIs • Debugging — production incidents traced to root cause, not patched over I am the one on the hook when my own products break, so I write integrations the way I want to maintain them: explicit error handling, real logging, and no silent failure modes. Happy to start with a short paid review before any larger commitment — you get a written assessment of what is fragile and what it is costing you, and we both find out whether we work well together.