Notes from
production.

Designing voice gateways that don't lie

The hard part of a voice agent isn't the language model. It's the seven layers of latency, hand-off, and silence detection sitting between an incoming SIP packet and a coherent response — and the user-experience consequences of getting any of them subtly wrong.

A short field guide to building gateways that fail honestly: when to barge-in, when to back off, why turn-taking is the new prompt engineering, and how to keep your eval harness honest about the parts your users actually notice.

Evaluating RAG: what to measure, what to ignore

Half the RAG metrics you see in dashboards are downstream of the same failure mode: confident retrieval over the wrong chunks. This post walks through the eval rig I default to — groundedness, faithfulness, answerability, refusal quality — and how to actually wire them into a CI pipeline that catches regressions before they ship.

Agentic systems and the boring trade-offs nobody mentions

Agent frameworks have gotten good at the demo. The interesting part — and the part that determines whether your production agent is a durable feature or a liability — is the boring stuff: tool authorization, conversation state, recovery on partial failure, and the operational debt of an autonomous process you can't interrupt.

From monolith to event-driven, without the rewrite

The two-year rewrite is almost always wrong. Notes from migrating the Storable Auctions platform from a synchronous monolith into an event-driven architecture — incrementally, while keeping the business running and the bidders happy.