skydiscover-ai/skydiscover

AI-Driven Scientific, Algorithmic, and Systems Discovery

What it solves

SkyDiscover addresses the difficulty of manually designing high-performance algorithms and complex software systems. It automates the discovery of better algorithms and the creation of specialized, end-to-end systems tailored to specific hardware, workloads, and requirements, replacing one-size-fits-all software solutions.

How it works

The framework operates through two primary modules:

  1. SkyDiscover-Optimize: Uses evolutionary search algorithms (such as AdaEvolve and EvoX) to iteratively improve programs. Users provide a scoring function, and the AI evolves the code to maximize that score.
  2. SkyDiscover-Synthesize: Employs coding agents to build complete systems from plain-text descriptions or formal specifications. It uses two main approaches: a formal-proof-driven method that generates code and machine-checked proofs together, and a test-driven method where agents iterate on implementation while an auditor generates new tests to prevent "reward hacking."

Who it’s for

It is designed for researchers and engineers in computer science and systems programming who need to optimize algorithmic performance or build high-efficiency specialized systems (e.g., key-value stores, LLM inference engines, or model routers).

Highlights

  • Adaptive Search: Includes AdaEvolve (which tunes its own settings) and EvoX (which optimizes its own optimization process).
  • Just-in-Time Systems: Capable of synthesizing systems that outperform general-purpose alternatives, such as KV stores faster than Redis and inference engines with higher throughput than vLLM.
  • Verification: Supports both formal machine-checked proofs and rigorous test-driven auditing for reliability.
  • Agent Integration: Works with various coding agents including Claude Code, Codex, Cursor, and Pi.

Related

  • Project
    SakanaAI/ShinkaEvolve

    A framework that combines LLMs and evolutionary algorithms to automate the discovery and optimization of scientific code across various programming languages.

  • Project
    algorithmicsuperintelligence/openevolve

    OpenEvolve is an open‑source Python library that combines large‑language‑model prompting with a MAP‑Elites evolutionary algorithm (island parallelism) to automatically rewrite, optimize, or invent code across many languages. Users supply an initial program and an evaluator; the system mutates the code via LLMs, evaluates candidates, and maintains a diverse Pareto‑optimal archive. It supports multiple LLM back‑ends (OpenAI, Gemini, Claude, local Ollama, etc.), offers deterministic reproducibility, extensible hooks, and can be run via CLI, Docker, or as a Python module. Example use‑cases include GPU kernel tuning, novel algorithm discovery (e.g., adaptive sorting, circle packing), scientific computing, and competitive‑programming solution generation.

  • Project
    argilla-io/distilabel

    A framework for synthesizing synthetic data and AI feedback to create high-quality datasets for training and fine-tuning AI models.

  • Project
    Yu-Yang-Li/StarWhisper

    A comprehensive AI suite for astronomy that provides specialized LLMs, light curve and spectral analysis tools, and a framework for autonomous telescope observations.

  • Project
    NVIDIA-NeMo/DataDesigner

    A framework for generating high-quality synthetic multimodal datasets with support for field dependencies, validation, and integration with external tools.