EVOLYTH: the Engine of Discovery
Autonomous AI research system for evolving machine-learning model architectures
Research in data science heavy domains is still bottlenecked by human iteration. We reflect about what experiments to run, how to inspect logs and metrics, compare charts, and decide what to try next. That process is slow. It is biased by what we already know. And it leaves too many promising ideas unexplored.
The real opportunity is faster scientific search: turning architecture design into a repeatable evolutionary loop.
Evolyth is an autonomous architecture evolution engine: a system that proposes model mutations, safely rewrites code, runs experiments, judges the results, and decides what should evolve next. This is a system for building the system that discovers better models.
Built around the tiny-cifar demo, Evolyth demonstrates measurable architecture improvement while preserving full lineage, reproducibility, leaderboards, and Pareto-front analysis. Sumbitted for Claude Code Hackathon 2026 organized by Anthropic.
See the full demo here