// writing

NOTES & BUILD LOGS.

Things I've learned shipping AI & products to production.

Cutting LLM inference cost 60% with semantic caching

An embedding-keyed Redis cache that skips the model on near-duplicate questions — and how to keep it safe.

Workload Identity Federation end to end: how an AWS Lambda calls Vertex AI with zero stored credentials.

Marketing site, learning platform, AI chatbot, video studio, data pipelines, and infra — shipped by one person.

Code-enforced guards that let an agent read and learn a live enterprise UI without ever mutating data.

Schema retrieval, few-shot assembly, and validation loops that make natural-language SQL actually reliable.