Claimly
An insurance-grade vehicle damage estimator. Upload one photo of a car and Claimly identifies the visible parts, spots the damage, and works out a priced, claim-ready repair estimate for that make and model.
Front bumper
Dent, lamp broken
Replace · R8,320
Overview
Motor claims usually start with a photo and end with a human deciding what the repair should cost. Claimly compresses that first mile. A client or assessor submits a single image; two computer-vision models run on it locally: one segments every visible panel, the other finds the damage, and the two are matched so each dent, scratch or shattered light is attributed to the part it actually landed on.
From there it is priced. Costs resolve per make and model in South African Rand from a catalogue the business controls, the assessor confirms repair or replace on each part with the total updating live, and the finished quote exports as a branded PDF. Detection is genuine and reviewable, never a stock diagram, with a person always in the loop.
What it does
Reads the actual photo
Parts and damage are outlined on the image itself. Hover any part to see the model's confidence and what it found.
Damage attributed to parts
Both models' masks are rasterised onto the image and overlap-measured, so a panel covered by damage is flagged and typed.
Priced by make & model
A managed catalogue resolves replacement and repair costs per vehicle in Rand. Switch the car and the estimate reprices on the spot.
Repair or replace
Human-in-the-loop review. The assessor makes the call on each part and marks anything the model missed; the total follows.
Branded PDF export
One click produces a clean quote with the photo embedded and a damaged-parts table, ready for the claim file.
Runs on your machine
Inference is CPU-only and local. Photos are analysed on the box; nothing is sent away.
How it works
Send in a photo
The client or assessor uploads a single photo of the vehicle. Add the make and model if known, so the estimate uses the right vehicle and pricing.
Let the models assess it
One model identifies the visible panels. Another detects damage. The two are matched to show exactly which parts were affected.
Review and send it on
Choose repair or replace for each affected part, watch the estimate update as you go, then export a clean PDF ready for the claim.
Under the hood
- Two pretrained YOLO11 instance-segmentation models run on the uploaded image: a parts model (23 classes: bumpers, doors, glass, lights, mirrors, hood, tailgate, trunk, wheel) and a damage model (6 classes: crack, dent, glass shatter, lamp broken, scratch, tyre flat). Class names are read from the models, so swapping weights needs no code change.
- Damage-to-part attribution works by rasterising both models' masks onto the original image canvas and measuring overlap ratio. A part covered by damage is flagged, and its damage types drive a repair-vs-replace recommendation.
- Pricing is a first-class data model: a canonical Part holds base costs, a PartPrice overrides them per VehicleModel, and each estimate snapshots its detected parts with polygon, confidence, damage and a frozen price, so later catalogue edits never rewrite an issued quote.
- Live, no-reload editing: JSON endpoints reprice the estimate when the vehicle or a repair/replace toggle changes, driven by a small vanilla-JS overlay. The photo and its SVG overlay share the image's pixel viewBox, so regions stay aligned at any size.
- Models load lazily and cache process-wide, so only the first request after a restart pays the load cost. Inference is CPU-only, a few seconds per image with no GPU required. The branded PDF is generated with ReportLab.
- Python
- Django
- Ultralytics YOLO11
- PyTorch
- NumPy
- Pillow
- ReportLab
- SQLite
- JavaScript
- HTML / CSS