Ornith
Ornith AI develops self-scaffolding large language models that iteratively improve their own code generation and optimization capabilities. Its GrandCode system ranked first in three consecutive Codeforces live competitions, outperforming all human participants and leading AI models. The company also open-sources CUDA-L1, a reinforcement learning framework achieving 3.12x average speedups across 250 KernelBench kernels.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Writing high-performance GPU code and competitive programming solutions requires deep technical expertise and manual, time-consuming optimization. Existing AI systems for code generation lag behind expert human performance, and automated tools for CUDA kernel optimization fail to match hand-tuned libraries like cuBLAS. Developing efficient implementations for tasks like matrix multiplication and nearest-neighbor search remains a bottleneck.
Solution
Ornith AI builds self-scaffolding large language models that iteratively improve their own code generation and optimization capabilities. The company's flagship GrandCode system beat every human participant and leading AI models in three consecutive Codeforces live competitions, establishing a new performance bar for AI in competitive programming. Ornith AI publishes its research and open-sources code for its systems, including CUDA-L1, a reinforcement learning framework that optimizes CUDA kernels using execution speed as the sole reward, and CRINN, which applies contrastive reinforcement learning to approximate nearest-neighbor search. The CUDA-L1 system trains on A100 GPUs, achieving substantial speedups over strong baselines and generalizing across GPU architectures. This research demonstrates a pathway toward autonomous software engineering and performance-critical code generation.
Target Audience
Primary customers are AI research labs, high-performance computing teams, and software engineering organizations seeking state-of-the-art automated code generation and GPU optimization capabilities.
Features
- GrandCode system that achieved first-place rankings in Codeforces Rounds 1087, 1088, and 1089, outperforming all human and AI competitors
- Self-scaffolding LLM architecture where models generate, evaluate, and iteratively refine their own code and scaffolding
- CUDA-L1 reinforcement learning framework for automated CUDA kernel optimization, achieving 3.12x average and up to 120x peak speedups across 250 KernelBench kernels
- CRINN technique that frames ANN algorithm design as a reinforcement learning problem, using execution speed as reward with accuracy constraints
- Research on evaluation-hacking defenses to prevent self-evolving systems from gaming performance metrics