Platform
How the damage assessment actually works
No brains, no neural glow. Four steps, each one checked against what carriers actually wrote on real claims.
Step one: identify the vehicle
The VIN decodes to year, make, model and where possible the submodel. The shop history is checked first because it carries the submodel that the public database usually leaves blank, and the submodel is what decides which bumper cover the car actually has.
Step two: read the damage
A vision model reads the photographs and proposes affected panels and operations. It is told what carriers actually decide on each panel, how many refinish hours they allow, and which removal and reinstall lines they pay to reach the damage, so the first write includes the access work an estimator usually adds by hand. When the damage cannot be seen clearly it asks for a wide angle of the whole accident instead of guessing, and any line the model is unsure of is dropped before you see it.
Step three: ground it against settled work
Proposed parts are checked against the comp index built from carrier-settled documents: 1,933 comparison lines drawn from 4,046 carrier-paid part rows. Where a part number matches, the price is what a carrier already paid for that exact part. Coverage by make is published on the parts page.
Step four: reprice at the right rate
The same hours and parts are priced across several labor rate profiles, because a repair written at a shop retail rate and the same repair written at a carrier rate are different numbers for the same work. Every estimate carries all of them, so you see the gap before the adjuster does.
What the price stands on
Parts are grounded in OEM part information and carrier-settled comp lines, and that is what the estimate stands on. A marketplace price fills in only where the model produced none, needs at least two new-condition listings that agree, and is overridden by a carrier comp wherever one exists.
Before you ask
Which model does it use?
Vision and text models from commercial providers, with the estimating logic, the comp grounding and the repricing built here. The model is a component, not the product.
What is it grounded on?
Carrier estimates of record. 1,933 carrier-paid comp lines drawn from 232 parsed carrier documents across 14 named insurers, plus measured repair-or-replace rates over 1,012 panel decisions and measured refinish hours. Those go into the prompt on every estimate, so the number you get is anchored to what carriers actually paid rather than to what a model guessed. That is the product. The corpus is aggregate: no VIN and no claim number is in it.
Is the model trained on my estimates?
No, and nothing is trained at all. No model is fine-tuned here and nothing is uploaded to a provider for tuning. The grounding above is carrier paperwork, not customer work, and your estimates are not published, sold, used as examples, or fed into any index. A shop that buys this does not have its work turned into someone else’s training set.
What happens when it is wrong?
You correct the line. Every line is editable, and the correction is the point: a partly right first write you can edit beats a blank screen.
Run it on damage you have already written
It is free to start: four estimates a month, with no card and no invite code. Bring your own photographs and a repair order you have already closed, so you can see how close it lands before you decide.