Flint — Live Meeting Co-Pilot

Real-time AI co-pilot desktop app for high-stakes live conversations — client consulting calls, discovery meetings, negotiations, and other professional dialogue where you need private, in-the-moment guidance. Captures remote participant audio from the call, transcribes locally with Whisper and RNNoise, and fires parallel directional, depth, and clarifying LLM threads in a stealth overlay invisible to screen-share. RAG over session context via sqlite-vec keeps guidance grounded in your brief, notes, and domain material. Phase 1 ships the job-interview domain (mock sessions, rehearsal, preferred answers) as the first vertical.

AI/LLM Ongoing 2 months
View Code
Flint — Live Meeting Co-Pilot

Technology Stack

Rust Tokio Tauri 2 React 18 TypeScript Tailwind CSS Whisper RNNoise VAD RAG sqlite-vec fastembed Groq Ollama Supabase

Key Results

Live meeting co-pilot for consulting and professional calls; Phase 1 = interview vertical
Product Scope
Local Whisper — zero audio leaves the device
Transcription
Groq cloud with Ollama fallback; sub-second P95 response time
Inference
sqlite-vec + bge-small-en-v1.5 — fully on-device RAG
Context Store
Overlay excluded from screen-share capture via Tauri OS compositor hints
Stealth
3 LLM threads (directional, depth, clarifying) fired simultaneously per question
Parallel Threads
API keys stored in OS keychain only — never on disk or in env
Keychain

Challenges & Solutions

  • Capturing system audio without touching the microphone to isolate the remote participant in live calls
  • Running Whisper transcription and RNNoise suppression in real time with sub-second latency
  • Firing parallel directional, depth, and clarifying LLM threads per detected question without blocking the UI
  • Building a stealth overlay with Tauri that stays on top, is transparent, and is excluded from screen-capture APIs
  • Local RAG over session context using sqlite-vec and bge-small-en-v1.5 embeddings fully on device
  • Designing a domain-agnostic conversation engine whose first shipped vertical is job interviews
  • Graceful failover from Groq cloud inference to a local Ollama model when the network is unavailable
  • Storing API keys only in the OS keychain — never in config files or environment variables

Project Stats

1
Team Member
2
Months
Back to Projects