All
How Ruby Voice holds a conversation at 240 milliseconds
Frontier reasoning and telephone-grade latency pull in opposite directions. Here is the architecture we used to stop choosing between them.
Teaching a voice model to be interrupted
Barge-in is treated as an audio problem. Most of the failures we found were conversational, and they were not fixable in the audio pipeline.
What OpsBench found in 40,000 simulated calls
Assistants that act can fail in ways a question-answering benchmark cannot see. We built an evaluation for the failures that cost money.
Routing across frontier models without changing your code
Heyno One sends each request to whichever model actually suits it. The interesting problem was not routing: it was knowing when the route was wrong.
When an assistant should stop and ask
Approval prompts are usually tuned by intuition. We tried tuning them by reversibility instead, and most of our assumptions did not survive.
Running a frontier assistant with the internet unplugged
Everything an assistant does casually in the cloud becomes an engineering problem on one machine in a basement. Here is what it took.