Majlis

Multi-LLM discussion arena: several models answer in parallel, you moderate with kicks, Curator moves (ready debate prompts), and optional web search (Tavily). History is persisted per room.

What you get

Feature	Notes
Curator	Recommends panelists; synthesizes after each surface/discuss round; shallow “same answer twice” check on depth
Curator moves	GET /curator/discuss-prompts — e.g. Challenge, Find consensus, Devil’s advocate — UI sends mode: "discuss" with a structured directive
Discuss round	Models only see active panelists in the pasted context — kicked models are excluded from follow-up discussion
Tools	Per-participant web_search at join; force_web_search on /discuss runs one shared Tavily pass for that round
UI	Light shadcn/Tailwind; stacked response cards; Clerk auth

Architecture

Layer	Tech
Frontend	React + Vite + TS, Tailwind, shadcn/ui → Cloudflare Pages
Backend	FastAPI → AMD droplet (Docker + Nginx)
Models	vLLM (ROCm) — parallel 8B-class workers + curator endpoint
Auth	Clerk (JWT verified with JWKS)
Search	Tavily
DB	SQLite + SQLModel

Project layout

majlis/
├── backend/           # FastAPI — routers/rooms, discuss, curator, feedback
├── frontend/          # Vite app — Pages, hooks/useDiscussion, components/ui (shadcn)
├── docker-compose.yml # vLLM × N + backend + nginx (droplet)
├── nginx.conf
├── docs/
│   └── MAJLIS_SLIDES.html   # Print / Save as PDF (browser: Print → PDF)
└── .env.example

Quick API reference

Method	Path	Purpose
POST	/room/create	Create room + curator recommendations
GET	/rooms	List user’s recent rooms
GET	/room/{id}	Room state
GET	/room/{id}/messages	Full transcript (model_id, display_name, layers)
POST	/discuss	SSE stream — body: message, optional target_participant_id, mode, force_web_search
GET	/curator/discuss-prompts	Curator moves (labels + instructions)
GET	/curator/models	Models + reputation for category
POST	/room/{id}/participant/add	Add model (web_search, thinking, fast_mode)
POST	/room/{id}/participant/{pid}/dismiss	Kick + reason string

Deploy (summary)

1. Droplet: docker compose up -d with .env (HF_TOKEN, TAVILY_API_KEY, CLERK_*, FRONTEND_ORIGIN).
2. DNS: Point api. and app host to droplet / Pages (Cloudflare proxy OK).
3. Frontend: VITE_API_URL, VITE_CLERK_PUBLISHABLE_KEY → npm run build → Pages deploy (or GitHub Actions in .github/workflows/).

Note: GitHub runners may be blocked from SSH to your droplet; backend updates can be rsync + docker cp + restart, or open SSH to Actions IPs.

Local dev

cd backend && pip install -r requirements.txt && uvicorn main:app --reload
cd frontend && npm i && npm run dev

Slides (PDF)

docs/MAJLIS_SLIDES.pdf — 16-slide deck (thesis, product depth, architecture, API sketch, security, trade-offs, roadmap, demo script). Source: docs/MAJLIS_SLIDES.html.

Regenerate: open the HTML → Print → Save as PDF (set margins to Default or Minimum so @page in the file controls the box). Or:

google-chrome --headless --no-pdf-header-footer --print-to-pdf=docs/MAJLIS_SLIDES.pdf file://$(pwd)/docs/MAJLIS_SLIDES.html

Environment

See .env.example for HF_TOKEN, TAVILY_API_KEY, CLERK_SECRET_KEY, FRONTEND_ORIGIN, model URLs, etc.

Models (typical)

Key	Role
llama-3.1-8b, qwen2.5-7b, mistral-7b, deepseek-r1-8b	Panel
curator	Routing, warnings, synthesis, discuss-prompt copy

Ports are wired in docker-compose.yml / config.py.