Kernel Sense is an autonomous 4-agent AI Lab, built by Team Reason AI System, that predicts GPU kernel performance using AMD compute to optimize AMD compute. Tuning GPU kernels traditionally means brute-force benchmarking every candidate configuration on real hardware — slow and expensive. Kernel Sense trains a model that predicts a configuration's runtime without running it, so only the most promising candidates need to be benchmarked live, cutting most of that brute-force cost out of the workflow entirely. Unlike a fixed AutoML pipeline that hands you a black-box model and a score, Kernel Sense is built so every agent justifies its output, not just produces one. Four LangGraph agents — Analyst, ML Engineer, Critic, Chief Scientist — collaborate autonomously: the Analyst explains why it made each preprocessing decision, the Critic actively judges whether the winning model is trustworthy and can reject its own pipeline's result and trigger a genuine retry loop, and the Chief Scientist synthesizes a business-ready deployment verdict grounded in real computed metrics rather than a templated summary.
Category tags: