Work · Client · the firm · 2026
AI operations for a foundation's assistant
Reliability, observability and cost work on a production AI assistant serving a large foundation.
What was built
- Model-performance monitoring and cost optimisation — routing, caching, model choice
- An embedding-model migration
- A read-only benchmarking harness against production data
- A source-cited cost and strategy brief for leadership
What happened
Brief delivered to leadership; a silent context-compaction defect diagnosed by reading the code, not the vendor's claims.
What it taught me
Verify what a system writes, not just what it reads back.