Notes & essays

What I learned cutting a small model's prompt nearly in half while making it more accurate, by treating the prompt like an API instead of a wish.

A denoiser that was 2.5× faster looked like an easy win, until I measured the thing that actually mattered. A short note on benchmarking what you care about.

A mid-call audio dropout that had nothing to do with audio, and what it taught me about debugging real-time systems on shared infrastructure.

A rolling deploy was killing in-flight queries during the drain window. The fix was one line of YAML, but only once I understood how the pieces shut down.

Swapping a managed cloud message queue for a self-hosted one inside the cluster cut latency and cost, but it handed us back a few things the cloud was quietly doing for us.

Call records come in a dozen flavours, each with different fields and lifecycle. A registry pattern keeps the decoder open to new types, and turns the one ugly exception into a documented invariant instead of a special case.

A real-time voice pipeline written in Python hits a ceiling at moderate concurrency. Rewriting it in Go bought roughly 10× the throughput per machine, but the real work was the benchmark that told the truth.