README.md

metadata

title: Ken AI Co-Listener
emoji: 🎧
colorFrom: yellow
colorTo: green
sdk: docker
app_port: 7860
pinned: false

Ken — AI Co-Listener for Professional Conversations

An AI co-listener that gives explainable, personalized interventions during professional conversations — built on AMD ROCm.

Quick Start

pip install -r requirements.txt
python server.py

Open http://localhost:7860

Environment Variables

export AMD_ENDPOINT="http://your-endpoint:8000/v1"
export MODEL_NAME="Qwen/Qwen3-14B"

How It Works

1. Onboarding — User selects a domain (Legal/Medical/Immigration/Career) and describes their situation
2. Processing — Upload a conversation recording → Whisper transcribes → LLM detects 4 trigger types
3. Session — Cards appear time-synced as the video plays

Trigger Types

Trigger	Detects	User's Gap
Jargon	Domain-specific terminology	"What does that mean?"
Impact	Content affecting YOUR situation	"How does this affect me?"
Question	Vague/hedge language	"What should I ask?"
Tracked	Dates, amounts, action items	"Will I remember this?"

Tech Stack

• ASR: faster-whisper (base model, CPU)
• LLM: Qwen3-14B on AMD Instinct MI300X via vLLM + ROCm
• Backend: Flask
• Frontend: Vanilla HTML/CSS/JS (Ken design system)
• Infrastructure: AMD Developer Cloud

Project Structure

server.py          — Flask app (run this)
pipeline.py        — Whisper transcription + parallel LLM trigger detection
triggers.py        — 4 prompt templates
templates/
  index.html       — Onboarding page
  session.html     — Session/processing page

License

MIT

Built With

• AMD Instinct MI300X (192 GB HBM3)
• AMD Developer Cloud + ROCm
• Qwen3 (Alibaba Cloud)
• vLLM inference engine