Sentry: AI Yield Copilot for Semiconductor Fabs

Semiconductor fabs often find a broken tool weeks after it breaks. An etch chamber's edge ring wears, nothing alarms, and every 25-wafer lot it processes is damaged until wafer sort catches it, often around 12 weeks later. At roughly $20,000 per advanced-node wafer, each of those lots puts about $500,000 at risk. Sentry closes that gap. Every wafer map has a shape, and the shape points to the cause: an edge ring suggests the etch or thermal chamber, a scratch suggests the handling robot or CMP polisher. Sentry classifies each map into one of 9 WM-811K patterns with a calibrated 3-CNN ensemble (0.912 macro-F1, 9.5 ms per wafer on CPU, scored on lots it never trained on). When one tool produces the same pattern three times in 24 hours, it raises an alert showing the evidence wafers, the likely cause with its source, the machine to check and the dollars exposed. Ask Sentry, an open LLM agent served on AMD GPUs through AMD Token Factory, answers questions about the line in plain English using 11 tools over live line data. Uncertain wafers (1.1%) go to a person, and the engineer decides whether to hold lots, open a ticket or dismiss, with every decision logged. In our replay of 500 real wafers through four simulated etch tools, Sentry flagged the drifting tool on its 3rd Edge-Ring wafer with zero false alarms beforehand. Built with PyTorch, FastAPI, Docker and Hugging Face Spaces.