Catch

Inspiration

You forget someone's name five seconds after the handshake. You promise "I'll send you that tonight" and it's gone by the elevator. Someone cuts in, you answer them, and your second point is lost. You can't glance at your phone mid-conversation without being rude. But the G2 sits in your line of sight, and its microphones sit at the perfect spot to pick up these problems. We wanted to embody quiet tech.

What it does

It catches names, promises that are made, and if you are interrupted, it prompts you with your last words to help you never forget any thread that a conversation might go. On top of that, it journals conversations that you deem important, storing the promises and a summary (of blurbs) of your conversation highlighting key terms and not preserving word-by-word transcription.

How we built it

We used Even Hub plugin (TypeScript, Vite, Even Hub SDK): streams the G2's microphone to our server, together with Even's per-frame speaker role, and draws what the server sends back. The same page doubles as Catch's phone screen in the Even app.

Server (TypeScript on Node, SQLite, hosted on Fly.io):
Deepgram nova-3 for streaming speech-to-text, with smart formatting and name keyterms.
Gemini 3.5 Flash-Lite for detecting promises (inside three seconds) and writing the jogger, in two passes: a draft, then a tidy pass checked so it never adds anything that wasn't said.
Spec first:
Before the event we wrote Catch's behaviour as Gherkin feature files and a log of design decisions using a custom (open-source) agentic behavioral driven development workflow.
During the event, every phase went through a review pipeline: AI coding agents built each piece, two independent reviewers checked it, a head reviewer deliberately broke the code to prove the tests caught it, and nothing merged until the full suite passed.
Tests never touch the network: speech arrives as timed words on a virtual clock, and model calls replay recorded answers. All build work was prompted into Claude Code. ## Challenges we ran into Unanticipated contradictions in our planning led to last-minute compromises and decisions to cut certain features. ## Accomplishments that we're proud of Making code that connects AI with the physical world and our physical senses. ## What we learned Hardware is not easy, but it is rewarding and we are proud of our project. ## What's next for Catch Ironing out lasting issues, and working on making the product more minimalistic and user friendly.