A self-hosted multi-agent framework that runs entirely on local LLMs. Agents reason over large data inventories and call deterministic tools for retrieval and verification. When a result fails a check, a bounded self-correction loop feeds the failure back and retries. Every output is typed and fully auditable. No data leaves the environment for a third-party API.
Built on Pydantic AI, the system coordinates several local LLM agents through a typed pipeline. Each agent shares a typed dependency context, emits Pydantic-validated structured output, and calls deterministic Python tools, so the model proposes and code verifies every result. The design targets auditability and governance: each step is traceable, and the final output is auditable for human review. I built it to test how far self-hosted models can drive a real, end-to-end agentic workflow without a cloud API, and benchmarked several local model families on the same task.
Coordinates multiple local LLMs (via Ollama) through a typed Pydantic AI pipeline, with no hosted API and no data leaving the environment.
Every agent returns Pydantic-validated structured data, and deterministic Python checks audit each claim. The LLM labels, the code audits.
When an output fails a check, the verifier feedback is fed back to the agent and the step retries, so results converge to ones that pass. Every final recommendation stays traceable for human review.