A real-time AI debate training platform with live transcription, Gemini-powered coaching, emotional analysis, ELO ranking, and global analytics.

• Structured rounds — Opening (3 min) → AI Opening → Rebuttal (2 min) → AI Rebuttal → Closing (1 min) → AI Closing
• Topic + side selection — Choose from preset topics or enter your own; argue For or Against
• Difficulty levels — Easy / Medium / Hard / Expert (each maps to an ELO bracket for matchmaking)
• AI opponent — Gemini 2.5 Flash generates contextual speeches, played via ElevenLabs TTS

• Browser captures mic audio via MediaRecorder and streams binary chunks over WebSocket to the backend
• Backend accumulates the full audio buffer and sends it to ElevenLabs REST STT every ~3s for interim results
• Final transcription triggered on round end

• After each user speech turn, Gemini analyzes the full transcript and returns 3 coaching tips (good / suggestion / warning)
• Per-entry feedback annotations shown inline in the live transcript
• Intervention monitor — polls every 10 seconds for Point of Order (POO) or Point of Information (POI) opportunities; shows an animated alert banner (POO = red, POI = amber)

• Records 4-second audio clips every 5 seconds during a debate
• Sends to Hume Expression Measurement streaming API (prosody model)
• Maps Hume's 48 raw emotions to 4 display categories: Confidence, Conviction, Calmness, Engagement
• Real-time WPM calculation using delta word count between ticks

• Semantic evidence search against a ChromaDB vector store
• Pre-seeded with 20 debate evidence entries covering AI regulation, UBI, nuclear energy, social media, crypto, space, parliamentary procedure, and more
• Upload custom evidence files or ingest text snippets in-session
• Shown in the right panel during debates for quick evidence lookup

• Webcam captured at 5fps, mirrored via canvas, streamed over WebSocket to backend /stream
• Visual analytics (gesture, eye contact, posture scores) accumulated during debate and shown on the results page

• Dual AI summaries — Gemini 2.5 Flash and K2 Think V2 (MBZUAI IFM) analyze the full transcript independently and show side-by-side assessments
• K2-specific features — logical fallacy detection, chain-of-thought reasoning breakdown
• Argument flow — visual map of how arguments developed and were rebutted across the debate
• Segment-by-segment review — each user speech turn scored across dimensions (Argumentation, Evidence, Delivery, Rebuttal, Structure) with individual strengths and improvements
• Detailed review summary — radar chart + score rings across all dimensions, plus a final overall score
• ELO update — displayed as a badge (+/- ELO) with new rating shown immediately after the debate
• Voice analytics — WPM pacing bar with target zone (130–160 WPM) and Hume emotional profile
• Full transcript — collapsible, with inline Gemini feedback per entry

• Standard ELO formula with K=32
• Win threshold: score ≥ 60
• AI opponent ELO by difficulty: Easy=1000, Medium=1400, Hard=1800, Expert=2200
• Rank tiers: Beginner → Intermediate → Advanced → Expert → Master → Grandmaster

• All users ranked by ELO descending
• Shows name, ELO, wins, losses, and tier badge

• Personal stats: ELO, wins, losses, win rate, rank progress bar
• Recent debate history with scores and ELO changes

• Full history list per user
• Individual debate detail pages

• Embedded Hex.tech dashboard at /global-stats
• Shows win rate vs difficulty, logical fallacy frequency, Hume emotional scores, and ELO progression across all users
• Data exported via GET /api/history/export

• Python 3.9
• Node.js 18+
• ffmpeg (required for Hume audio conversion)

brew install ffmpeg

cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Create backend/.env:

ELEVENLABS_API_KEY=...
GEMINI_API_KEY=...
HUME_API_KEY=...
DATABASE_URL=sqlite:///./dialex.db
CHROMA_PERSIST_DIR=./chroma_db
ALLOWED_ORIGINS=http://localhost:3000

Start the server:

uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

API docs: http://localhost:8000/docs

cd frontend
npm install

Create frontend/.env.local:

NEXT_PUBLIC_BACKEND_URL=http://localhost:8000
NEXT_PUBLIC_API_URL=http://localhost:8000
NEXT_PUBLIC_HEX_APP_URL=https://app.hex.tech/019d3775-c0a1-700d-8635-53b17531662a/app/032qCqqyoiPDwMpBeipGqh/latest

Start the dev server:

npm run dev

App: http://localhost:3000

• Gemini quota — Free tier exhausts quickly; 429 errors reset daily
• ElevenLabs — Uses REST API (not WebSocket streaming); WebSocket returns 403 on Creator plan
• Hume — Requires HUME_API_KEY; ffmpeg must be installed for audio conversion
• RAG — Auto-seeds on first debate start; ChromaDB persists to backend/chroma_db/ (gitignored)
• Python 3.9 — Avoid X | Y union syntax; use Optional[X] / List[X] from typing