Skip to content

Evolyth

Evolyth is an autonomous AI research loop that improves machine-learning models through repeated mutation, execution, evaluation, and review.

Code Video Cerebral Valley

Core trick

Coding agent proposes architecture edits, runs each training job on cloud GPUs, and a reviewer agent queues the next experiments.

Tags

dev toolsagents

Description

Evolyth is an autonomous AI research loop that improves machine-learning models through repeated mutation, execution, evaluation, and review. A Claude-powered coding agent proposes safe architecture changes, the system runs training jobs locally or on Google Cloud Run GPUs, then a reviewer agent analyzes the result and queues the next promising experiments. The project includes a live dashboard showing evolution progress, leaderboard, queue state, AI R&D cost, and total evolution time. It is designed to make model experimentation faster, more observable, and more reliable.
akaliutau/evolyth★ 20 forksPythonMITCreated Apr 25, 2026 · last push Apr 28, 2026

README excerpt

<p align="center">
  <img src="docs/main_screen.png" alt="Evolyth" width="680">
</p>

<h1 align="center">Evolyth</h1>

<p align="center">
  <strong>Autonomous, explainable AI R&amp;D for model architecture evolution.</strong>

<p align="center">
 <a><img alt="Status" src="https://img.shields.io/badge/status-mvp-orange"></a>
 <a><img alt="Python" src="https://img.shields.io/badge/Python-3.12%2B-blue"></a>
 <a><img src="https://img.shields.io/badge/Cloud-Google%20Cloud-4285F4" alt="Google Cloud"></a>
 <a><img src="https://img.shields.io/badge/Claude-Code%202.5-yellow" alt="Gemini 2.5"></a>
 <a><img alt="License" src="https://img.shields.io/badge/license-MIT-lightgrey"></a>

</p>




<p align="center">
  Evolyth turns a machine-learning research problem into an evolutionary loop:
  propose a mutation, run the experiment, review the result, preserve the evidence, and queue the next best move.
</p>

<p align="center">
  <a href="#quickstart">Quickstart</a> ·
  <a href="#why-evolyth-exists">Why</a> ·
  <a href="#system-architecture">How it works</a> ·
  <a href="#research-problem-contract">RP contract</a> ·
  <a href="#cloud-gpu-execution">Cloud GPU</a> ·
  <a href="#live-dashboard">Dashboard</a>
</p>

---

## [Why Evolyth exists](#why-evolyth-exists)

Modern AI coding agents can write surprisingly good experimental code. The missing piece is not code generation; it is **scientific discipline**.

A model-improvement loop becomes useful only when each step is:

- **bounded**: the ag

Team