A multi-agent reasoning system that runs any question through a 4-stage cognitive pipeline: psychological diagnosis → debate routing → parallel persona generation → synthesis.
Live: depth-chi.vercel.app (frontend) / Render (backend)
Four advisors with distinct, grounded philosophies respond to your question simultaneously, then a synthesis stage produces a 150-word consensus with concrete action steps.
| Advisor | Lens |
|---|---|
| Marcus | Risk and asymmetry (Taleb, skin in the game) |
| Alex | Strategy and leverage (Thiel, zero-to-one thinking) |
| Maya | Customer reality (Mom Test, what users actually want) |
| Turing | Engineering clarity (Brooks, what to simplify or cut) |
The same question, different panel: Worried Mom, The Hater, The Conspiracist, The Hype Man. Shorter responses, entirely different energy.
User question
│
▼
Stage 1 — Psychological Brief
Diagnoses the hidden fear behind the surface question.
Is this Validation Seeking? Fear of Failure? Identity Crisis?
│
▼
Stage 2 — Debate Router
Decides who speaks first, urgency level (1–10),
and the core tension the council should argue about.
│
▼
Stage 3 — Parallel Persona Generation
All 4 advisors run concurrently via ThreadPoolExecutor.
Each reads from a markdown knowledge base (Taleb, Mom Test,
Brooks, Thiel/Helmer) before responding.
│
▼
Stage 4 — Synthesis
A diplomatic summary acknowledging where advisors disagreed,
extracting what each was right about, and producing 3
time-bound action steps + 1 named pitfall.
- Backend: Flask + Groq (Llama 3.3 70B), deployed on Render
- Frontend: Vanilla HTML/CSS/JS, deployed on Vercel
- Concurrency:
ThreadPoolExecutorfor parallel LLM calls - Knowledge bases: Markdown files per advisor (strategy, risk, engineering, customer research)
cd backend
pip install -r requirements.txt
# Create .env
echo "GROQ_API_KEY=your_key" > .env
python app.py
# Backend at http://localhost:5000
# In another terminal, open frontend/index.html directly
# or use: python -m http.server 3000 --directory ../frontendAPI endpoints:
POST /council/debate— full 4-stage pipeline, returns all stages as JSONPOST /api/getResponses— Roast Council, fast parallel responsesGET /health— API + persona status check
The original idea was more ambitious: a full council of AI agents each with persistent memory, disagreeing with each other across sessions. That hit API rate limits quickly and the UX was chaotic.
What shipped instead is more focused — one question, one session, four disciplined perspectives, one synthesis. The psychological brief stage is the part that actually changed how it felt to use: instead of getting four opinions on what you asked, you get four opinions on what you meant to ask.