03 · pilot
Valuation Engine
A machine-learning pricing engine that reports the worse but honest number.
99.6k
market listings modeled
16.5%
average pricing error, honestly measured (was 18.4%)
80%
price-range coverage, as designed
52k
lines of code across the stack
/ interactive · pick a car, read the waterfall
try it, everything responds
adjustment waterfall · 2019 Ford Ranger 2.0 Bi-Turbo Wildtrak 4×4 A/T · 96,000 km
80% range
R408k–R452k
comps
41 listings
Mock cars, real anatomy: the production engine samples comparable listings from ~99.6k scraped rows, regresses the baseline to the car's mileage, applies these line items, then lets an XGBoost residual model correct where the rules are systematically wrong, and the confidence lens decides how much the model gets to speak. The interval is conformalized: 80% target coverage, 80% achieved.
A note on this demo. Every figure, name and data point here is invented for illustration. No client data, real economics or production configuration appears anywhere on this site. What is real is the machinery: the calculations, the rules and the logic running in your browser are the same ones the production system uses, so the demo behaves exactly as the real thing behaves.
What it is
/ in plain terms
A car dealer types in a vehicle and the system estimates what it is worth as a trade-in, along with how confident it is in that estimate. When we discovered that our accuracy score was flattering itself, we rebuilt the measurement so it could not cheat, published the worse number, and required every improvement to beat it honestly.
AI trade-in valuation for South African car dealers. Three layers, blended by confidence: a transparent rule engine, a market model trained on 99.6k scraped listings, and a residual model fitted against real dealer trades. The defining decision: the original 10.6% error figure turned out to be inflated by leakage, meaning the model had been accidentally tested on cars it had effectively already seen. It was rebuilt on an honest protocol that grouped every test by car identity and measured 12.3%, and the rebuild had to beat the champion on that harder number before it shipped.
Founder-engineer. Scrapers, machine-learning lab, serving engine, multi-tenant SaaS.
The load-bearing details
- 01Three-layer engine: a mileage-adjusted comparable-sales regression with line-item adjustments, a market model working in log space, a residual model that learns where the rules are systematically wrong, and a confidence-weighted blend of all three.
- 02Honest cross-validation: test cars are grouped by identity so the model can never be graded on a car it trained on, target encodings are rebuilt inside every fold, results are checked across three random seeds, and synthetic labels are quarantined. The leak hunting is documented rather than hidden.
- 03Calibrated uncertainty: the engine quotes a price range designed to contain the true price 80% of the time, and in testing it does exactly that.
- 04The whole engine was ported to TypeScript, including a hand-written evaluator for the trained model that matches the Python original bit for bit.
- 05A learning loop turns every confirmed trade into a training label. Retraining gates on the honest error measure, and every valuation is stamped with the model version that priced it.
- 06Auction hammer prices used as reduced-weight proxy labels, kept out of evaluation, cut the error on cheap cars from 21.2% to 18.7%.