In the Cleanest Rooms on Earth, an AI Watches for Exposed Skin A human hair is about 70 micrometers wide. A particle that can ruin a semiconductor wafer is 0.1 micrometers — 700 times smaller. In an ISO Class 1 cleanroom, fewer than 12 particles of that size are allowed per cubic meter of air. Every person who enters one of these rooms is a contamination bomb. We shed roughly 40,000 skin cells per hour. Each breath releases thousands of particles. A single exposed patch of skin between a glove and a sleeve can release enough particles to scrap an entire wafer lot — which, depending on the chip, can be worth anywhere from $10,000 to $200,000. That's why cleanroom workers wear full-body "bunny suits" — sealed coveralls, hoods, face shields, double gloves, and booties. And that's why a semiconductor fabrication company in Manisa asked me to build a system that detects gowning violations in real-time. ## Why Standard PPE Detection Doesn't Work Here I walked into this project thinking it would be straightforward. I'd built PPE detection for construction sites — hardhats, vests, masks. The models were good. I'd just adapt them. I was wrong. Construction PPE detection works by identifying the presence of specific objects (helmet, vest) on a detected person. The person looks like a normal human, and the PPE items have distinctive shapes and colors. In a cleanroom, the person doesn't look like a normal human. They're completely encased in white fabric. There's no visible face, hair, or skin. The body geometry is distorted — the hood adds volume around the head, the coverall obscures the waist and joints. Standard person detectors trained on COCO or OpenImages have significantly reduced accuracy on fully gowned individuals. And the violations aren't missing objects. They're gaps — a 2cm strip of exposed wrist between glove and sleeve, a slight opening at the collar where the hood doesn't overlap the coverall, a face shield that isn't fully sealed to the hood. ## Reframing the Problem Instead of "detect missing PPE items," I reframed it as "detect exposed skin or gaps in the gown envelope." This meant I needed a model that could: 1. Detect a gowned person reliably (not trivial given the unusual appearance) 2. Segment the gown boundary 3. Identify any region within the person's bounding box that shows skin, hair, or gap I trained a two-stage pipeline. Stage one: a YOLOv8 detector fine-tuned on cleanroom footage to detect gowned persons. Stage two: a lightweight segmentation model (based on a MobileNetV3 backbone) that classifies pixels within the detected person as "gown," "skin," or "gap." The training data was the hardest part. Violations are rare and intentionally corrected fast — nobody wants to be the person who contaminated a wafer run. I ended up staging violations with cooperating cleanroom technicians. We systematically created every type of gap: unsealed collars, short gloves, rolled sleeves, lifted face shields, improperly secured booties. ```python # Gap detection alert logic def check_gown_integrity(segmentation_mask, person_bbox): x1, y1, x2, y2 = person_bbox person_region = segmentation_mask[y1:y2, x1:x2] total_pixels = person_region.size skin_pixels = (person_region == SKIN_CLASS).sum() gap_pixels = (person_region == GAP_CLASS).sum() exposure_ratio = (skin_pixels + gap_pixels) / total_pixels # Any exposure above 0.5% triggers alert if exposure_ratio > 0.005: # Locate the gap region for the alert gap_mask = (person_region == SKIN_CLASS) | (person_region == GAP_CLASS) contours = cv2.findContours(gap_mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) return True, contours, exposure_ratio return False, None, exposure_ratio ``` ## Edge Deployment Constraints Cleanrooms have strict rules about what electronics can enter. No fans (they generate particles). Limited cabling (outgassing from PVC insulation). Everything must be wiped down with IPA before entering. We used fanless industrial PCs with passive cooling, mounted outside the cleanroom and connected to cameras inside via sealed conduit. The cameras themselves were cleanroom-rated models with stainless steel housings — not cheap. Each processing unit handles four camera feeds simultaneously. The model runs on TensorRT-optimized INT8, achieving 22ms inference per frame per camera. We process every third frame — at 30fps input, that's 10 effective checks per second per camera, more than enough for the walking pace of cleanroom workers. ## The First Real Catch Three weeks after deployment, the system flagged a technician in Fab 2 whose left glove had pulled back about 3cm from the coverall sleeve. The gap was visible for about 12 seconds as the technician reached up to adjust equipment on an overhead tool. The alert went to the floor supervisor's tablet within 2 seconds. The technician was asked to re-gown. A 12-second exposure in a non-critical zone — probably harmless. But in front of an open wafer cassette, it could have been catastrophic. The facility manager told me later that they'd had a contamination event six months earlier that was eventually traced to a gowning failure. That single event had scrapped 14 wafers and cost over $80,000 in direct losses, not counting the yield investigation and production delays. The entire AI system cost less than that one incident. ## What I Didn't Expect The system revealed behavioral patterns that nobody had documented before. Gowning violations are not random. They cluster: At shift changes, when workers are rushing. Violation rate increases 3x in the first 15 minutes of each shift. At overhead reaching positions, where sleeves ride up. Certain tools that required arm extension generated 60% of all glove-gap alerts. In warmer months, when workers are more likely to loosen collars. June through August showed a 40% increase in hood-seal violations. This data didn't just catch violations — it changed how the facility designed its gowning process. They added wrist-securing elastic bands to gloves, redesigned the collar seal on coveralls, and adjusted tool heights to reduce overhead reaching. ## The Broader Lesson PPE detection is one of those problems that sounds solved until you actually try to deploy it in a specific domain. Construction, food processing, semiconductor fabs, pharmaceutical clean areas — each environment has fundamentally different requirements for what "compliant" means, what violations look like, and what the consequences of failure are. The value isn't just in catching violations. It's in generating data about when, where, and why violations occur. That's what enables systemic improvement, and that's what makes the difference between a camera that watches and a system that actually makes things safer.