AMPM | Devpost

Meeting AI Agent, Parrot
AMPM Web Application

Inspiration

You pick up a task: "Implement rate limiting on auth." Why auth? What threshold? The ticket doesn't say. You search Slack—47 results. You ping the PM—they're in meetings. An hour later, you still don't have context. So you guess.

This is the SDLC Memory Problem. As work moves from meeting → PRD → code, context dies. Engineers waste 5-10 hours/week hunting for "why."

What it does

AMPM is an AI agent that joins your meetings and becomes your team's memory.

In-Meeting Q&A: Ask "Why did we choose OAuth?" mid-call and get an answer in 2 seconds with sources
Knowledge Graph: Connects meetings → decisions → action items → code
Ripple Detection: Change a requirement and instantly see all downstream impacts before anyone builds the wrong thing

How we built it

Cerebras (Llama 3.3 70B) — Ultra-fast inference (~400ms) makes real-time responses possible
NetworkX + FAISS — Knowledge graph with vector search
Whisper + ElevenLabs — Speech-to-text and natural voice responses
Playwright — Browser automation for Google Meet
Streamlit — Web dashboard

Challenges we ran into

Audio routing into Google Meet — Had to implement virtual audio (BlackHole/VB-Cable) to pipe bot output as mic input
Real-time performance — At 30s nobody waits; at 2s it's useful. Optimized entire pipeline for sub-3s responses
Interruption handling — Dual-thread architecture for listening while speaking

Accomplishments that we're proud of

Sub-3-second responses in live meetings
Full meeting-to-code traceability
Ripple detection that prevents mistakes before they happen
Natural conversational interface with wake words and interrupts

What we learned

Speed is a feature. 30-second vs 2-second responses isn't UX polish—it's the difference between useful and unused. Fast inference unlocked this entire use case.

Knowledge graphs > vector search alone. "What's affected if we change X?" requires graph traversal, not just semantic similarity.