01The photograph
4032 × 3024 · lot lighting · handheld
The same vehicle imagery a buyer sees on the listing, including imperfect angles, inconsistent lighting, reflections, and low-quality photographs.
Bucepha Intelligence Two models, and a ledger that corrects them.
One model reads the photographs and says where the damage is. Another prices it. Every estimate is stored beside the invoice that settled it, so the next estimate is answerable to the last one: the defect that was found, where it was on the car, how bad it was, and what it cost.
That join is the whole point. A model reading photographs alone can tell you a bumper is damaged. A model reading photographs that are linked to invoices can tell you what that damage costs, because it has seen what was actually paid.
- 4,012
- Images processed
- Every estimate
- Stored beside the invoice
- 1+
- Downloads, all three apps
Proprietary vehicle data
Not a generic vision API.
A pipeline of our own, priced the way dealership buyers and insurers actually evaluate vehicles.
The difference is the data. An estimate starts from a costed labor guide and is corrected by what this dealership’s own repairs actually came to.
Published openly, and honest about which part stays private.
Three things read a photograph, and they are not the same kind of thing. Our computer vision model finds where the damage is on the panel. A pretrained language model describes what it is and how bad. Our own records of what the repair actually cost decide what it is worth.
Our detection model
Localization of visible damage, published on Hugging Face.
huggingface.co/buckets/shalin-code/bucepha-cardamage-detectionOur dataset, never shared
The corrections dealerships make, joined to the invoice that settled the repair. It stays in our own storage and is not published, sold, or used to train anybody else's model.
Pretrained language model
Reads each photograph and writes what it sees, in the words a body shop prices from.
The two models read the same photograph independently and are compared afterwards. Agreement makes a finding stronger. Disagreement is kept rather than resolved, and the report says so, because a confident answer nobody checked is worth less to a buyer than an honest doubt.
Ground truthSample record, illustrative
VIN ..M12345 · 12 photos
Dealership lot, as photographedPHOTO 04 / 12
PHOTODEFECTSEVERITYREPAIRCOST
- Marker
- 01
- Component
- Front bumper, left
- Defect
- Repaint with orange peel texture
- Severity
- 4 / 5
- Repair
- Repaint
- Invoice
- $480
Training support
1,284 similar repair records
More linked repairs behind a defect means a tighter estimate. Fewer means the model reports lower confidence rather than committing to a number.
The model does not simply learn what damage looks like. It learns what that damage means financially.
One vehicle. One ground-truth record.
01
Photo
The same image quality a buyer sees.
02
Defect
Identified and located by a dealership professional.
03
Repair
What the dealership actually decided to fix.
04
Invoice
What the repair actually cost.
05
Estimate
Answerable to the whole record.
A defect without its repair history is only half the answer.
A general vision model can identify a car, a wheel, or a damaged bumper. It does not know that a particular bumper was resprayed, whether the finish is acceptable, or what that repair actually cost in a specific market. We train on the complete record.
Generic vision
CarBumperWheelDamage
What will it cost?
Guess
Vehicle defect intelligence
PhotoLabeled defectSeverityRepair orderInvoice
What will it cost?
Learned from real repairs
A pipeline of our own, built for the way dealership buyers see cars.
Every training example connects what the vehicle looked like, what was wrong with it, and what it actually took to put it right.
02The mark
010203
3 marks · severity 1 to 5
Every defect is placed exactly where it appears and assigned a severity from one to five. The annotation becomes the same numbered reference the product uses when reporting the vehicle.
03The invoice
Repair order, illustrative
| Bumper refinish | $480 |
| Wheel repair | $140 |
| Tire replacement | $120 |
| Total | $740 |
The repair invoice connects the visual defect to the actual work performed and the actual amount paid. Line by line.
Photo
Annotation
Repair order
Invoice
Model training example
The model learns the relationship between the defect and the bill.
- 01Orange peel
- 02Bumper respray
- 03Body shop
- 04$480
Training supportHigh support
1,284 similar repair records
- 01Kerbed rim
- 02Wheel repair
- 03Tire replacement
- 04$260
Training supportThin support
17 similar repair records
More historical examples behind a defect produce a tighter estimate. Where the record is thin, the model widens the range and says so rather than implying a precision it does not have. Counts shown are illustrative.
Repair costs are not theoretical. They are local.
Repair economics vary by market, repair facility, vehicle, part, severity, and historical repair behaviour. The same bumper respray is not the same number in two cities, so the pricing side of the model prices from where the work was actually billed rather than from a national average.
Estimated cost, bumper respray by market
Illustrative
- Northeast metro$540
- Midwest$470
- Southeast$430
- Mountain$505
- West coast$580
Learned from historical repair transactions rather than scraped generic estimates.
How the data is labelled
A model is only as good as what it was shown.
Annotation workstationExample 04 / 12
01Annotation
- Defect type
- Repaint
- Location
- Front bumper, left
- Severity
- 4 / 5
- Confidence
- High
- Component
- Bumper cover
- Repair category
- Refinish
Ground-truth record
- Defect
- Front bumper
- Severity
- 4
- Repair
- Repaint
- Invoice
- $480
Reviewed
- Human reviewer
- Label verified
- Invoice matched
Every label has an answer behind it.
Generic dataset
ImageAnnotation
This platform
Image
Defect
Repair
Invoice
Outcome
Ground truth is not what someone thinks the repair should cost. Ground truth is what the repair actually cost.
The data advantage compounds
More linked examples, not more images.
A larger pile of photographs teaches a model what a bumper is. A larger pile of records that join visual condition to the repair that followed and the amount it cost teaches it what a bumper is worth. Those are different assets, and only one of them gets harder to copy over time.
- 1 repairComponent
- 10 repairsComponentDefect
- 100 repairsComponentDefectSeverity
- 1,000 repairsMakeModelComponentDefectSeverity
- 10,000+ repairsMakeModelYearComponentDefectSeverityMarket
Illustrative. Each row adds the dimensions a larger linked record set can separate.
Build vehicle AI on real repair data.
Turn dealership and insurance records into a vehicle intelligence system priced from the defects, repairs, and costs that matter.