Project · 2026
← All workLive Context Collision
Real-time meeting agent that surfaces conflicting organizational memory while a decision is still being made: a prior legal block, a capacity conflict, a decision the company already reversed.
2nd in two Vultr categories + DeepMind prize, Milan AI Week
- Speechmatics
- Gemini
- Next.js
- TypeScript
- Vultr
Overview
Live Context Collision listens to a meeting in real time and compares each new statement, promise, or timeline against a shared company memory graph. When it finds a collision, it shows an evidence-backed warning card during the call. Specialist Gemini judges propose collisions and deterministic code verifies them, so every card quotes the memory fact verbatim. Built at the Milan AI Week hackathon (AI Olympics, 2026).
The problem
- Catching a conflicting fact while the decision is still open, before it lands in the notes
- Making every warning card quote real evidence
- Keeping latency low enough for live use
What we built
- Speechmatics real-time transcription with speaker diarization feeds each finalized utterance into local retrieval over the memory graph
- Five parallel Gemini specialist judges propose collisions; a deterministic verifier rewrites the quote to the memory fact or drops the card
- Service-agent actions run on a queue off the hot path; the personal summary is generated after the meeting
Results
- 2nd place in two Vultr categories and the Google DeepMind special-track prize at Milan AI Week (AI Olympics, 2026)
- Detects five collision types: legal/compliance, previous decision, priority/capacity, dependency blocker, customer promise
- 220 passing Vitest tests plus Playwright end-to-end specs in the repository
How it works
Speechmatics transcribes the meeting live and tags who is speaking. Each finalized utterance goes through local lexical retrieval over a company memory graph: seed facts plus facts learned from earlier meetings.
Five Gemini specialist judges run in parallel, one per collision type: legal or compliance, a previous decision, priority or capacity, a dependency blocker, a customer promise. A deterministic verifier then rewrites each card’s quote to the memory fact’s exact text, or drops the card. No second model sits on the hot path.
Service-agent actions go to a BullMQ queue, off the hot path. After the meeting, a curator writes what was decided back into memory and a personal summary is generated. It is built to deploy to Vultr in EU and US regions behind a Cloudflare Worker that routes by geography, so EU data stays in the EU.