
Automotive vision/damage evidence
REVYR
A working browser proof that geometrically registers issue and return photos, locates new damage proposals and verifiably transfers each proposal to a human.
Working vehicle damage proofThe case in 30 seconds
REVYR
- Problem
- A working browser proof that geometrically registers issue and return photos, locates new damage proposals and verifiably transfers each proposal to a human.
- Solution
- From issue and return photo to registration, local damage detection, independent verification and human inspection.
- Role and focus
- Damage vision, registration
- Status
- Working vehicle damage proof
- Result
- Platform, demo, evidence suite
Live lab / try it yourself
Compare issue with return
Using two vehicle photos, check the geometric alignment and assess which new damage proposals deserve human investigation.
Frozen test suite
Four ports, one fixed case
The proof measures reproducibility within one controlled image pair. The results are convincing for that suite, but deliberately not a fleet-wide accuracy claim.
Dent and scratch localized, without additional notification.
Photometric variants deal exactly zero new damage.
Perspective, rotation and translation remain within the pre-selected limit.
Two label-free proposals independently confirmed ten times.
Geometry before conclusion
Place the same panels on top of each other
Differences in location, perspective and cropping may otherwise appear as apparent damage. Drag to compare issue and registered return.


Chain of custody
No magical look
REVYRseparates recording quality, geometric registration, local model proposal, independent verification and human decision.
Test photo first
Angle, sharpness, opacity and light determine whether comparing is useful.
Align panels
Vehicle anchors bring the same parts into one comparison space.
Local AI highlights
The ONNX model represents class, score and region without evaluator or target coordinates.
Man decides
A confirmed region remains a concern, not an automatic claim or liability.

40/40 is strong and limited
The current suite proves that this fixed case reproducibly passes through damage, zero, registration and verification gates. The next credible step is a blinded fleet pilot with other vehicles, angles, weather conditions and damage sizes.
- 30 unique local entries with hash and pixel checking
- Raw inference saved before scoring
- Evaluator not included in the model
- No independent claim or liability decision
Open the platform
From photo to inspection
View the complete product site, run the fixed demo or access the reproducible evidence file with the test ports and limits.
More concept work
Other working proofs
From care vision and security evidence to voice operations.
Care vision / human verification
Security vision / provability
Voice agents / business operations
AI / logistics / operations