The Sound of Nothing: when transcription AI invents a conversation from silence
Our whole product is built on remembering what was said. So the most dangerous thing we can ship is a confident memory of something that never happened — and we found one hiding two characters past every guard we had.
Read →Is this the same person? Resolving identity with a deterministic net and an LLM receptionist
A shared email doesn't mean the same human — families share everything. A new shape for one of software's oldest problems: deterministic recall, judgment in the middle, and a system engineered so the model's worst day can't do real harm.
Read →Never lose a second: a recording pipeline that survives crashes, dead zones, and force-quits
You get one take — the color's already processing, the lamp's already curing. Here's how we make sure the audio always comes home, even when the phone, the network, and the user are all working against us.
Read →Auditioning a voice: choosing a speech-to-text engine like it was a hiring decision
Word error rate told us almost nothing. The questions that mattered were the ones a model can pass while still being wrong in exactly the way that hurts a memory product.
Read →The wait that wouldn't scale: getting your notes the instant you stop, no matter how long the appointment
The longer the session, the longer you waited — backwards for a product built on memory. So we stopped doing the work after you stop and started doing it while you work. A 60%+ cut to the wait, and the win grows with length.
Read →