AI · Technical PM
Maetri: evidence-qualified AI
AI designed to represent what it observed, what it inferred, and how confident it should be.
Evidence
Target
up to 65%
Design target · Ontario LTC study protocol targets ≥35%
Projected
$41,180
Savings per prevented pressure-ulcer event · CMS/AHRQ citation required
Target
κ ≥ .85
Validation protocol target · not completed validation
Modelled
$62K/mo
Modelled: $10M annual billing × 15% denial rate × half never recovered ≈ $750K a year · assumed inputs, not measured
Stakes
In care settings, a record can survive while its meaning is lost. An AI that reads video and writes a confident note is dangerous when it cannot say how sure it is.
The bet
Build AI that keeps what it saw, what it inferred and how confident it is as separate things — and hands the judgment to a clinician.
Three moves
Observe, interpret, validate, record. Keyframes from a public caregiver-training video were tagged to clinical frameworks (Omaha System, Braden) with a claim, a confidence and a rationale for each. Why: every statement stays traceable to the frame it came from.
Built in a stop. Where the image could not support a judgment, the pipeline said so — “cannot assess from image” — instead of guessing. Why: knowing what you don’t know is the product.
Defined validation before claiming it. A protocol with an AI–nurse agreement target and nurse acceptance criteria, written before any claim of accuracy. Why: demonstrated is not validated.
What changed
Demonstrated on one run (December 2025) using a public training video and synthetic patient data — no patient information. The validation protocol is defined; validation is not yet complete.
What I’d do next
Run the validation protocol in partner care homes and publish the agreement results, whatever they are.