Case studies(2)
Production engineering work with the real numbers: what was broken, what I changed, and what it did for users.
Cutting the wait out of AI workflows
The median run went 340s → 300s and the slowest tenth dropped 87 seconds, without touching a single model or prompt.
- Typical run
- 40s sooner
- typical run
Making AI generations fail less often
A reliability layer took generation failures from about 12 in every 100 down to fewer than 1, across 962,376 scored generations.
- Failures per 100 generations
- 12 → under 1
- per 100 generations
Sanitized by design. Product, repository, and PR names are removed on purpose. Every number on this page is real and was measured in production. Argue with the method below.