// op_013 — case study In Dev
AI Wingman
// the problem
High-stakes conversations — a sales call, a pitch, a first date, a hard negotiation — are where advice matters least in retrospect and most in the moment. Every prep tool and conversation notes app operates before or after the fact; nothing coaches you while you're actually talking, when the next sentence is the one that counts.
// architecture decisions
Live Transcription
Deepgram streams live transcription of the conversation audio, giving the coaching layer a real-time transcript to reason over instead of a recording processed after the call ends.
Real-Time Coaching
Claude runs over a persistent WebSocket connection, reasoning on the live transcript stream to whisper the next move through AirPods with low enough latency to actually be useful mid-conversation.
Conversation Memory
PostgreSQL persistence tracks names, interests, promises, and objections across sessions — a follow-up conversation next week starts with the context of what was already said, not a blank slate.
Background Reliability
A background audio session keeps coaching running with the screen locked, and the app surfaces mic and server health honestly instead of silently failing mid-conversation.
// key outcomes
The app is organized around a simple loop — prepare, talk, review, follow up, improve — with dedicated Briefs, Practice (voice roleplay against Claude before the real conversation happens), History, Playbooks, and Text Coach surfaces. Apple and Google Sign-In, PDF transcript export, OTA updates, and Sentry monitoring round out a platform built specifically for conversations where the stakes are real.
// related builds