AURA. Multi-agent voice copilot, inspired by Paris Fashion Week
- 7 (aura-fashion-ai repository) agents, 4 of them in the main cascade
- 2,189 (aura-fashion-ai repository) lines of Python
- 48 h solo, at a hackathon
Problem
Sabrina Ramonov’s hackathon, March 2026: 48 hours, solo. The topic I picked comes from Paris Fashion Week: a voice copilot for fashion showroom sales.
Decisions
- Split the work into agents: 7 in total, 4 of them in the main cascade.
- The cascade: an extractor turns the sales rep’s sentence into a record (buyer, store, item, quantity, price); a checker looks up stock and margin; a strategist decides to accept, counter, upsell, offer an alternative or suspend; a copywriter drafts the confirmation email once the deal is agreed.
- A voice pipeline: browser microphone, WebSocket, Deepgram for transcription, the agents on Groq, Cartesia to answer out loud.
Result
A complete prototype in 48 hours, published on GitHub.
What’s not done yet
- AURA was never deployed.
- No automated test suite.
- Demo data only: fictional buyers and catalog.