Demo · May 2026 · Deen Developers' Vibe + Tell, London

AutoScout, live

Name: AutoScout, live
Start: 2026-05-01
Location: Deen Developers' Vibe + Tell

My demo at Deen Developers' Vibe + Tell, a show-and-tell night for things people have built with AI. AutoScout scores UK used cars using 33 million MOT tests. I presented it with a live deck that pulls real car data, so anyone in the audience could search any car.

AutoScout · How It Works. Scroll the slides, or click them and use the arrow keys.Open full screen ↗

Why AutoScout exists

My brother needed a used car for an 80-mile commute, five days a week. Pick wrong and it's breakdowns, fuel bills and a repair bill bigger than the car. We had no idea which model held up at high mileage, which year was the sweet spot, or what it would cost to run.

We're not the only ones. 7.64 million used cars changed hands in the UK in 2024, and almost all of them were bought on dealer anecdotes, forum gossip and gut feel.

Auto Trader gets 90% of its marketplace revenue from dealers (£509m of £565m in FY25), so the dealers are its customers. It also assumes you already know which car you want. AutoScout takes zero money from dealers and covers the part before that: what car suits me, and which model and year.

One score from three engines

Every popular UK used car gets a Vehicle Suitability Score from 0 to 10, built from real data instead of anecdotes.

Reliability, 40%"Will it break?" DVSA MOT pass rates from 33 million tests, Parkers expert ratings, known faults, and UK and US recalls.
Running costs, 35%"What does a year cost?" Fuel is 65% of this one: official MPG, the real-world gap and live pump prices. Then servicing, insurance and tax.
Depreciation, 25%"Will I lose money?" Live AutoTrader prices for 559 cars against their list price, with a penalty for volatile EV values.

Why raw MOT data lies

Before scoring, AutoScout corrects for nine sources of human error. Old cars always fail more, so each car is scored against its age group: a 3-5 year old car is expected to pass 85% of the time, 6-9 years 75% and 10+ years 65%.

Cars under three years old don't need an MOT, so a 100% pass rate there is noise. Fewer than 5,000 tests caps a car at 80 out of 100. Failures are weighted too: a structural fault counts in full and a worn tyre counts for 0.2.

A £3k Fiesta failing on worn tyres and thin pads says more about the owner than the car, so those failures count half. And every one of the 3,420 models counts equally, so a Fiesta with 50,000 tests doesn't outweigh a Porsche 911 with 600.

AutoScout also won't rate a car it can't prove. Most Chinese EVs haven't been on UK roads long enough to know if they'll hold up, so they're not scored yet.

Inside the prompt

The model never recalls facts from memory. Before it writes a word, it gets eight feeds of verified data for that car and generation: price, official spec, insurance and tax, known faults, recalls, MOT pass rates, the manufacturer's reputation and the wider market. In plain English: we hand the AI a fact sheet and tell it to only use these numbers.

I ran the same prompt across three models on every car. Gemini 3 Flash was sharpest on UK trim details. Grok 4 Fast was a near-tie at a fraction of the cost per call, so that's what shipped. Gemini 2.5 Flash Lite was about 17% quicker but stayed generic. The anchored prompt mattered more than the model.

Fast is the product

Most visits never touch a model. About 65% of traffic is served instantly from pre-generated pages for the top cars, 12% from the browser's memory in about 150ms, and 18% from the Supabase cache in about 500ms. Only about 5% goes to a fresh Grok 4 Fast call through OpenRouter, which takes around 8 seconds, with every data feed fetched at the same time.

The stack is React and Vite on Vercel, Supabase, Upstash Redis, Cloudflare R2 for images, Gemini and Grok through OpenRouter, Browserbase for reading AutoTrader and Parkers, PostHog, Resend, and Playwright running 256 tests across 5 devices. Claude Code is my pair programmer.

1,000+ users in a day, then I listened

I posted AutoScout on Reddit, and 1,000+ people pulled the site apart in 24 hours. It got banned from several car subreddits for looking AI-built, and the criticism was often right. Every comment and DM went into the backlog.

Accuracy at scaleOne electric car slipped through with a petrol service schedule. One wrong fact and the whole site loses trust, even when the other 499 pages are right.
The data is the productI built v1 on static CSV files. When DVSA API access came through, I rewrote the whole reliability engine, and some "top picks" turned out to have real recalls.
Throw the first version awayI shipped a servicing tab in a day, and it was wrong from the start. I rebuilt it properly once the DVSA data came in.
Cheap images, big speed winWith no budget for a photo API, I wrote my own scraper and put every image behind Cloudflare's free CDN. It was the biggest speed win the app has had.
Check the cached pagesThe model sometimes drew depreciation graphs going up, or wrote copy that read like ChatGPT. Every page still needs a human pass before it goes live.

Prompts to copy

AutoScout gives the AI a fact sheet for each car and tells it to only use those numbers. Here's a simpler version for a 2018 Honda Civic. Swap in the facts for your own car.

The fact-sheet prompt

The first two lines are the rules. Everything in [brackets] is a fact you check yourself first.

car-guide-prompt.txt

You're writing an honest used car guide for UK drivers. // who it is for

Only use the facts below. Don't make up any numbers. // the main rule

If something isn't backed up by these facts, say so.

Car: [Honda Civic, 2016-2021, 5-door hatchback]

Typical price: [£13,999 on AutoTrader]

Fuel: [petrol, 54.3 mpg]

Insurance and tax: [group 15-18, £180 a year]

MOT pass rate: [89.2% for 3-5 year old cars]

Known problems: [oil thinning on short trips, air-con damage from stone chips]

Recalls: [Takata airbag, emissions software]

Write the guide. Every number must come from the facts above. // the task

Plain Text13 lines · UTF-8

Shipping fast only works if you're willing to throw the first version away. Today, a user finds a bug, I route it to Claude and the fix lands. Next, I want Claude to spot the issues from real usage and fix them without me in the middle.