Damage type
Chip · bullseye
Vision · Geometry · Decision
Open the camera.
Walk toward the damage.
Done.
An AI that reads a windscreen the way a customer already holds a phone. No forms. No reference cards. No coins. One walk in, one quote out, with the measurement done for you on the spot.
Live demo
01
Spotting and sizing are separate jobs.
The AI finds the damage on the screen. Geometry, not the AI, turns it into millimetres. We never ask the model to estimate size.
02
Two measurements must agree.
The same walk produces two independent readings of the damage, by completely different methods. If they don't match, we don't commit.
03
Cautious by default.
When the engine isn't sure, it sends the captured imagery to a human reviewer or leans toward replace. A wrong repair costs far more than a careful replace.
02 / 08 Walk to your windscreen
One walk in. One answer out.
Step 01 / 04
Open the camera
A customer holds their phone, walks toward the windscreen. No forms, no reference card.
01 · Approach
02 · Plate
03 · Measure
04 · Decision
Scroll to advance
We're already working on it the moment the camera opens.
From the very first frame, before the customer has even thought about framing the damage, the engine is reading the number plate, looking up the vehicle, and pulling the exact dimensions of that windscreen from the parts catalogue. By the time the user is close enough to capture the chip, half the work is already done.
Plate readDVLA lookupParts catalogue< 2s
{
"plate": "AG07 NGV",
"confidence": 0.984,
"vehicle": {
"make": "BMW",
"model": "3 Series · F30",
"year": 2018,
"colour": "Mineral Grey"
},
"windscreen": {
"width_mm": 1402,
"height_mm": 812,
"part_no": "BMW-43R-2117",
"heated": true
}
}
Frame index
042 / 187
Distance to vehicle
≈ 3.2 m
Plate height (px)
68 px
Two ways to measure.
One walk.
No card. No coins.
The customer didn't snap, didn't tap, didn't reposition. That same walk gives the engine two completely different ways to size the damage in millimetres, and a quote is only committed when both ways agree.
Method 1 · The screen is the ruler
○ STANDBY
Detected windscreen
1402 × 812 mm
Corner detection0.97
Scale at centre1.84 mm/px
Damage size (Method 1)
— mm
Damage outline74 × 81 px
SourceScreen geometry
Method 2 · The walk is the ruler
○ STANDBY
Frames analysed
187 frames · 0.6 m walk
Tracked points11,418
Reconstruction errorsub-pixel
Damage size (Method 2)
— mm
3D points on damage2,113
SourcePhone motion
Agreement check
○ PENDING
0mm
5mm
10mm
15mm
20mm
R1 · screen · 8.2mm
R2 · walk · 7.9mm
Δ = 0.3 mm · 3.7%AGREEMENT —
Size
8.0 mm
Distance to edge
11.4 cm
Driver line of sight
Outside
Agreement
96.3 %
Confidence
High
Repair
The rules: chip under 30 mm · more than 70 mm from the edge · not in the driver's line of sight. All three pass → repair. A 30-minute resin slot is offered. The rules made the call, not the AI.
Agreement commits a quote.
Disagreement gets a human check.
The two measurements of the same chip are compared. Within ~10–15 % of each other → a quote is committed on the spot. Further apart, or the damage is too small to size reliably → the captured imagery is sent to a trained reviewer. The system never bets on a borderline case.
The repair-vs-replace rules (size, distance from the edge, line of sight) are applied as plain rules, not by the AI. The AI handles pixels; geometry handles millimetres; the rule book handles the call.
When the engine isn't sure,
a trained human is.
The AI is allowed to commit a quote. It is not allowed to talk itself into one. Borderline cases (methods that disagree, damage too small to size, edges and lines of sight that need a judgement call) are routed to a small team of trained reviewers who see the same imagery the engine saw and confirm the call in seconds.
Reviewers don't re-do the work. They confirm, override, or request a re-capture. Every override flows back into the training set the next morning.
Median time on a review
18s
From the case appearing in the queue to the reviewer's decision being sent back to the customer.
Reviewer agreement with senior re-check
99.6%
Sampled weekly. The 0.4% disagreements become training material, not silent corrections.
Share of jobs that ever reach a human
8.1%
The other 91.9% are committed by the engine on the spot. The review team exists so that number can stay honest.
Reviewer console · LIVE
ARAisha R · senior reviewer
Queue3 waiting
Method 1 · screen9.4 mm
Method 2 · walk12.8 mm
Difference
3.4 mm · 31%
Engine call
Hold
Notes
Star break, partial occlusion from rain on the outer surface. Engine flagged the disagreement at 0.42s.
Your brand.
Our intelligence.
Same engine. Same two measurements. Same careful decision. Wrap it in your customer's brand: colours, type, chrome. Ship it under your name.
The platform is white-label by default. Every surface (the in-app camera, the on-screen overlay, the booking flow) is themed at runtime. Click a swatch to repaint it live.
Run a real walk through the engine.
Three pre-recorded walks: a repair, a replace, and a borderline case that gets routed to a survey. Each comes with the full pipeline output ready to play. Click one. Watch what the engine sees.
Or upload your own walk-in (≤ 8 MB / 5 s). Limited demo · production uses the full method. Rate-limited 5 / IP / hour.
Pipeline · Walk-in #1AG07 NGV · BMW 3 · 2018
PlateMethod 1Method 2Decision
Plate
—
Method 1 (screen)
—
Method 2 (walk)
—
Agreement
—
Verdict
—
Part
—
Engineering you can run a P&L on.
Measured on a separate set of 4,200 real walk-ins covering 38 vehicle models across UK and EU operating regions.
Verdict accuracy
94.2%
Repair vs. replace, agreement-committed cases · n = 3,118
Median latency
1.8s
From last frame to verdict on production fleet · p50
Sent for human review
8.1%
Cautious by default · prevents false-repairs
Form fields → photo
5 → 1
Customer-facing input collapsed to a single walk