Don’t Panic
Interview copilot (desktop)
A macOS interview copilot that captures both sides of a call, transcribes in real time, and generates live answer suggestions — with a private on-device STT option.
The challenge
Live interviews and calls move fast, and generic transcription tools blur who said what. The brief was a real-time copilot that separately hears both sides of a conversation and suggests answers in the moment — privately if the user needs it.
“Real-time, source-attributed answer suggestions during a live conversation.”
The approach
- 01
Built dual-source audio capture — “You” via mic and “Them” via a virtual/system-audio device — attributed separately.
- 02
Streamed both channels through Deepgram for real-time speech-to-text, with an offline Apple on-device option for privacy.
- 03
Fed the transcript to a pluggable LLM layer supporting Claude, GPT-4o, and Grok through one adapter.
- 04
Packaged it as a signed native macOS app with subscription billing wired for distribution.
Design iteration
A modern rebuild of a stale idea
Don’t Panic is a ground-up reimplementation of a three-year-stale interview-copilot concept, rebuilt on a modern Tauri plugin architecture. The major net-new feature over every reference project studied was dual-source capture — separately attributing the two speakers — which reshaped the entire audio and transcription pipeline.
What we built
Highlights of the work
Dual-source capture
Hears you and the other party separately for clean attribution.
Real-time transcription
Streaming STT with a private, on-device fallback.
Multi-LLM answers
Claude, GPT-4o, and Grok behind one pluggable adapter.
Signed desktop distribution
A notarized macOS build with subscription billing.
Capabilities demonstrated
Add these to your own project
Let’s talk
Have a project, or a product that could work harder?
Most projects begin with a short, no-pressure discovery call.