RecoveryTrek — Real-Time Video Inference | Punch
← ALL WORK / AI PROJECTS / RECOVERYTREK · PROOF
RecoveryTrek RECOVERYTREK · PROOF · COMPUTER VISION · RECOVERYTREK.COM ↗

A witness in every pocket.

How Punch built AI verification for Proof, RecoveryTrek's remote drug-testing platform — the camera witnesses the test, so a person doesn't have to be in the room.

COMPUTER VISION REAL-TIME INFERENCE MOBILE COMPLIANCE
PROOF · REMOTE TEST, VERIFIED FROM THE CLIENT PROOF REEL
PRODUCT WALKTHROUGH
HOW IT LEARNED TO WATCH

The labeling, and what it bought.

HAND-DRAWN GROUND TRUTH
Data labeling Manually marking items for precise AI identification.
LIVE ON A HOME WEBCAM
Detection Identifying objects using colored vector spaces.
BEYOND THE BOUNDING BOX

Then we taught it to read the strip.

Drawing vector spaces around test objects teaches a CNN what things are. For OralTox panels we went further — a separate model that reads what the colors mean, interpreting color-coded litmus strips through ordinary web cameras: uneven lighting, cheap optics, real homes.

THE STRIP READER · WEBCAM FOOTAGE
FIG. 01 — ORALTOX PANEL, READ LIVE · A SECOND MODEL INTERPRETS THE COLOR-CODED STRIPS, NOT JUST THE OBJECT
FROM THE BENCH · UNSTAGED

The gate itself: YOLO evaluation on SageMaker.

Labels are only the start. A screen capture from the actual workflow — YOLO evaluation running on AWS SageMaker, scoring each candidate model against the held-out set before anything our floor labeled is allowed near production inference.

YOLO EVAL · AWS SAGEMAKER · SCREEN CAPTURE
FIG. 02 — THE EVALUATION RUN, UNRETOUCHED · CANDIDATE MODELS SCORED ON SAGEMAKER BEFORE DEPLOYMENT
WHY REMOTE TESTING NEEDED A WITNESS

A test no one watches is a test no one trusts.

Court-ordered and recovery-program drug testing traditionally means travel, scheduling, and a human observer. RecoveryTrek's Proof platform replaces the trip with a phone — if the result can be trusted.

Punch built the verification layer: real-time inference confirms the right person is taking the right test the right way — flagging anything questionable for human review, with the evidence attached.

VERIFIED ✓ HUMAN REVIEW ON FLAGS
“Designed, built, and shipped end to end by Punch.”
0TRIPS TO A TESTING SITE
24/7TESTS VERIFIED, ANY HOUR, ANYWHERE
100%OF FLAGS GET HUMAN REVIEW, WITH EVIDENCE
WHAT WE BUILT
On-device identity & liveness check — confirms the right person is present before the test begins.
Real-time test-step inference — verifies the right test is run the right way, frame by frame.
Anomaly flagging with evidence — anything questionable is routed to a human reviewer, with the clip attached.
A labeled training dataset — 50,000+ annotated frames behind the verification model.
CASE STUDY recoverytrek case study cover Case study PDF IN THE NEWS RECOVERYTREK.COM · OFFICIAL RecoveryTrek — monitoring that holds. ↗ LINKEDIN · OFFICIAL RecoveryTrek on LinkedIn. ↗
Punch / Lab NotesISSUE 14
LATEST BUILD
THIS ISSUEInside the models we shipped this month.What we trained, what surprised us, and the eval that let it ship.
PUNCH · AI RESEARCH NEWSLETTERThe AI we’re building next — in your inbox.Research notes from the lab, the models we’re training, and the new projects we’re shipping. A few times a month — no spam, no fluff.FREE FOR SUBSCRIBERS · THE AI EVAL PLAYBOOK — SEE INSIDE ›
RECENT ISSUESPer-subscriber avatarsYOLO eval at scaleOffline product recognition· JOINED BY 2,400+ ENGINEERS & FOUNDERS
Building AI where trust is the product? Talk to the engineers who built this →
PREVIOUS — REDCLOUDREDCLOUDWave the phone. The shelf uploads itself. NEXT — NETFLIXNETFLIXThe trailer that knows what you watched.