Notes
Writing about systems
that had to work.
What the eval cases caught, where the latency actually went, and the decisions that got reversed. Mostly voice agents, retrieval and the deployment underneath them.
Keeping a conversation alive while the user navigates away
Chat with file sharing in a recruiting platform. The upload was the easy half. The half that took thought was that switching dashboard views destroyed the thread and every attachment in it.
- react
- redux
- s3
- architecture
Two coaches, one interview, and a race nobody could see
A HubSpot form becomes a Slack message someone can claim in one tap. Straightforward, until two people tap at the same moment and the system tells both of them yes.
- aws
- lambda
- dynamodb
- architecture
An assistant with write access has to ask first
Reads and writes are not the same risk. Giving an agent a send button means designing for the moment it misreads you.
- agents
- safety
- postgres
- architecture
The obvious voice agent architecture was the slower one
Splitting a voice agent into specialists behind a router is the standard advice. I built it, measured it against a single agent, and deleted it.
- voice
- architecture
- latency
- evals
