La Colonia replaces single-model AI responses with a council of LLMs that deliberate, rank each other's answers, and synthesize a final verdict. Models take on Spanish-language personas — El Profesor, La Abogada, El Patrón — grounding the experience in Latino culture.
Built on Andrej Karpathy's llm-council concept and extended into a full-stack app: FastAPI backend streaming results over SSE, React 19 frontend with a radial deliberation view, and support for 9 LLM providers (OpenRouter, Ollama, Groq, OpenAI, Anthropic, and more).
App in Action
How It Works
Deliberation
4–8 council models answer in parallel. Results stream back as they arrive via Server-Sent Events.
Peer Ranking
Each model anonymously scores all other responses. Aggregate rankings surface the best answers.
Synthesis
A designated Chairman model reads all answers and rankings, then delivers a single authoritative verdict.
Tech Stack
Backend
- FastAPI + uvicorn (port 8001)
- Async httpx — concurrent LLM requests
- SSE streaming to frontend
- 9 provider adapters (OpenRouter, Ollama, Groq, OpenAI, Anthropic, Google, Mistral, DeepSeek, custom)
- DuckDuckGo / Tavily / Brave web search
- JSON file storage — no database
- uv for package management
Frontend
- React 19 + Vite
- Radial deliberation view
- react-markdown for responses
- 3-mode execution: chat only, + ranking, full deliberation
- Configurable personas per model
- Responsive, CSS-native design
Highlights
Key Outcomes
- Real-time streaming: SSE delivers partial results as each model finishes — no waiting for the slowest model
- Provider agnostic: Mix cloud (OpenRouter, Groq) and local (Ollama) models in the same council
- Culturally grounded: Spanish-language personas and prompts designed for the Latino community
- Zero-database deploy: Conversations persist as JSON; runs anywhere with
uvandnpm