skydiscover-ai/skydiscover

What it solves

SkyDiscover is a framework designed for AI-driven scientific discovery and the end-to-end synthesis of high-performance systems. It addresses the challenge of manually designing complex algorithms or systems (like LLM inference engines or key-value stores) by using AI to iteratively discover and optimize solutions based on formal specifications or performance metrics.

How it works

The project is divided into two primary modules:

• SkyDiscover Optimize: A modular framework for algorithmic discovery. It uses evolutionary techniques and LLMs to incrementally improve solutions. It provides a unified interface to run and compare various search algorithms (such as AdaEvolve and EvoX) across hundreds of optimization tasks, supporting multiple evaluator formats including Python functions, containerized environments, and Harbor-format benchmarks.
• SkyDiscover Synthesize: A toolset for full system synthesis. It employs two main techniques: Inductive-Deductive Synthesis, which uses formal specifications in Rocq to generate implementations and proofs, and Test-Driven Synthesis, which uses a team of planning, coding, critiquing, and auditing agents to iterate on implementations based on natural language descriptions and tests.

Who it’s for

This framework is is intended for researchers and engineers working on algorithmic optimization, scientific discovery, and the automated synthesis of performant software systems.

Highlights

• SOTA Algorithms: Includes AdaEvolve (adaptive search) and EvoX (self-evolving strategy evolution).
• Broad Benchmark Support: Natively supports Harbor-format benchmarks and integrates with suites like AlgoTune and BigCodeBench.
• Flexible Evaluation: Supports Python-based, Docker-containerized, and Harbor-format evaluators.
• Agentic Synthesis: Uses a multi-agent approach for test-driven system synthesis.
• Live Monitoring: Includes a dashboard for real-time tracking of evolution, code diffs, and human-in-the-loop steering.