Yield Guardian: AI Root-Cause Agents for Fabs

Every chip fab fights the same costly question: when yield drops, why? Engineers spend hours digging through hundreds of sensor traces and thousands of wafer maps before they can act. Yield Guardian collapses that work into minutes. It runs on real production data: the SECOM line, with 590 sensors per unit, and WM-811K, 811,457 real wafer maps from 46,293 lots. Start a run from a real sample, pick a case, draw a defect on a real wafer layout, or upload your own fab data, and watch the line compute live, station by station. Monitoring scores every unit and checks every sensor against its control limits, with a 1.3% false-alarm rate on held-out data. A residual CNN reads each wafer map and names its defect pattern with 0.90 macro-F1 and 97.8% accuracy on lots it never saw in training. It also says "unknown pattern" when a map looks like nothing it learned, and finds the most similar past wafers. When an alarm fires, one click hands it to a team of AI agents running on AMD. Using real tools, not guesses, they gather the evidence: what drives the risk, which sensors moved together, how healthy each sensor is, which wafers look alike. They write a root-cause report in which every number links back to the evidence that produced it. The AI never decides. An engineer reviews the report, approves the actions, and a work order opens on the line, with every decision recorded in an audit log. Nothing is hidden. Every model's accuracy is published with its 95% confidence interval, and a savings calculator turns those measured rates into hours and money for your own fab. Switch the whole system between AMD and Claude in one click. Yield Guardian is live today: real data, evidence-linked answers, and people in charge.