Features
Generates optimized CUDA/Triton kernels from any model
32 parallel Coder+Judge agents with MAP-Elites evolution
Pattern RAG with 1,711 CUTLASS and 113 Triton patterns
Up to 5× speedup over torch.compile(max-autotune) on Llama-3.1-8B
100% numerical correctness verification
Automatic Tensor Core optimization (WMMA, TMA) for Hopper
Three optimization modes: --turbo, default, --quality
Dual output: Triton Python kernels or native CUDA C++
Interactive CLI with wizard and benchmark browser
Session management for tracking past optimizations
Supports HuggingFace IDs, KernelBench tasks, or custom files
Credit system: 1 credit per kernel, 1-2 for HF models
Powered by fine-tuned Nemotron 3 Nano 30B at 250K tok/s
Drop-in replacement: same API, zero code changes
Reverse-engineer causal mechanisms of AI models
Reveal internal structure and hidden representations
Detect performative chain-of-thought in LLMs
Identify confounders and debug model behavior
Validate whether models learned real clinical understanding
Trace unstable behaviors to brittle internal features
Reduce hallucinations via features as rewards
Accelerate materials discovery with self-correcting search
Control training precisely with less data and off-target effects
Support for LLMs, life sciences, and robotics/vision models
Harvest activations from trillion-parameter models
SOC 2 Type II certified security and compliance
Analyze latent policy structure in robotics models
Interpret genomic models like Evo 2
Discover novel biomarkers via model reverse-engineering
