clabsi-audit-insights

Dressing audits were being done by eye and written on paper, and I wanted to know whether a phone camera plus a browser could do part of that looking.

CLABSI Audit Insights is the most ambitious of these builds and also the one I have to describe most carefully. It has three parts: statistical process control charts with proper P-chart control limits, a social determinants layer that maps non-clinical risk against infection risk, and a photo audit tool that runs an image model in the browser through @huggingface/transformers, using onnx-community/mobilenetv4_conv_small. The inference is real and runs on the device; what is not real yet is the clinical judgement layered on top of it. Inside performMedicalAnalysis the dressing verdict is simulated MVP logic, not a validated medical model. There is also no backend at all — no migrations, no tables — so nothing that gets audited is saved.

• Me: "Can the browser look at the dressing photo?" → Yes, through an ONNX model running locally. No image leaves the phone.
• Me: "Then can it tell me the dressing is unsafe?" → No. Classifying an image is not the same as clinical judgement, and the code says so.
• Me: "Track compliance over time properly." → P-charts with control limits, so a bad week is distinguishable from a real shift.

The gap between running a model and being allowed to trust it. The photo pipeline works end to end; the decision it produces is placeholder logic, and calling that AI diagnosis would have been dishonest.

Inference in a browser is now easy. Validation is not. And a safety tool with no database is a demonstration, not a system of record.

• The clabsi-audit-insights Story

  Photo-first CLABSI audits

  Ebook · $9.00

  Read it