Engineering & Distilled Thoughts

Engineering Dec 12, 2025

The Pursuit of < 500ms Latency

When we started Vocalis, the standard GPT-4 response time was 3-4 seconds. In a voice conversation, that silence feels like an eternity. It breaks the illusion of presence.

We rebuilt our entire ingestion pipeline using Edge Websockets and optimistic audio buffering. Here is how we managed to get the "Time to First Byte" (TTFB) of audio down to sub-500ms...

Read Full Post ->
Product Nov 28, 2025

Why We Built "The Shark"

Most AI assistants are designed to be servile and polite. But growth doesn't happen in a comfort zone. We realized our users weren't improving their negotiation skills because the AI was too nice.

"The Shark" is designed to interrupt you. It challenges your premises. It simulates the stress of a real salary negotiation with a difficult hiring manager. It's not nice, but it works.

Read Full Post ->