Project 04 / Full-Stack · AI

La Colonia: A Multi-LLM Council

A 3-Stage Multi-LLM Deliberation Council

FastAPI React 19 Vite OpenRouter Ollama Python SSE Streaming Async httpx react-markdown uv

Overview

01 / 04

La Colonia runs a council of LLMs instead of a single model. The models deliberate, rank each other's answers, and synthesize one verdict. Each model takes on a Spanish-language persona, El Profesor, La Abogada, El Patrón, grounding the experience in Latino culture instead of a generic chat interface.

I built it on top of Andrej Karpathy's llm-council concept and extended it into a full-stack app: a FastAPI backend that streams results over SSE, a React 19 frontend with a radial deliberation view, and support for nine LLM providers including OpenRouter, Ollama, Groq, OpenAI, and Anthropic.

Walkthrough

02 / 04

La Colonia start page showing council members — Start page · council roster

Question sent to council — 01 Deliberation · models query in parallel

Peer ranking leaderboard — 02 Peer Review · models rank each other

Final verdict from El Patrón — 03 Synthesis · El Patrón delivers the verdict

Three-Stage Pipeline

03 / 04

Deliberation Four to eight council models answer in parallel. Results stream back as they arrive over Server-Sent Events, so nobody waits on the slowest model.
Peer Ranking Each model anonymously scores every other response. Aggregate rankings surface the strongest answers without a human in the loop.
Synthesis A designated Chairman model reads every answer and every ranking, then delivers one verdict.

Stack & Outcomes

04 / 04

The backend runs on FastAPI with async httpx firing requests at up to nine LLM providers in parallel, OpenRouter, Ollama, Groq, OpenAI, Anthropic, Google, Mistral, and DeepSeek, then streams results to the frontend over SSE as each model finishes. There's no database; conversations persist as JSON files, so the whole stack runs anywhere uv and npm exist.

The frontend is React 19 on Vite, rendering responses through react-markdown inside a radial view that shows every council member at once. Three execution modes let you run chat-only, add peer ranking, or run the full deliberation pipeline, with personas configurable per model.

LLM Providers Supported

Deliberation Stages

Database Required