I’ve been experimenting with AI agents lately and built something I call PolyAgents – a lightweight framework that lets multiple LLMs think together on a user prompt and then summarizes or merges their insights using a local semantic consensus model.
How it works:
You send a prompt.
Three different Gemini models (Flash 1.5, 2.0, 2.5) each generate their own reasoning paths.
Then, a 90MB local model (all-MiniLM-L6-v2) processes their answers and builds a shared consensus (currently a smart summary, soon a full synthesis of ideas).
Stack:
Dockerized setup
Backend: Gemini API, Postgres, Redis, Qdrant
Frontend: React + TypeScript + Vite
Entirely open-source
It started as a fun side project because I had a bit of coder's block — but it turned out way more interesting than expected. It’s meant to be hackable and easily extendable. Think of it as a playground for AI reasoning strategies.
Still rough around the edges, but curious what this community thinks. Would love any feedback, ideas, or collaborators.
I’ve been experimenting with AI agents lately and built something I call PolyAgents – a lightweight framework that lets multiple LLMs think together on a user prompt and then summarizes or merges their insights using a local semantic consensus model.
How it works:
You send a prompt. Three different Gemini models (Flash 1.5, 2.0, 2.5) each generate their own reasoning paths. Then, a 90MB local model (all-MiniLM-L6-v2) processes their answers and builds a shared consensus (currently a smart summary, soon a full synthesis of ideas).
Stack:
Dockerized setup Backend: Gemini API, Postgres, Redis, Qdrant Frontend: React + TypeScript + Vite Entirely open-source It started as a fun side project because I had a bit of coder's block — but it turned out way more interesting than expected. It’s meant to be hackable and easily extendable. Think of it as a playground for AI reasoning strategies.
Still rough around the edges, but curious what this community thinks. Would love any feedback, ideas, or collaborators.
GitHub repo:[https://github.com/Fenix46/PolyAgents] My email: [scarlataemanuele@gmail.com]