AI car mechanic: expert diagnosis from a £20 scanner

A friend's YouTube video showed how to disable the rev limiter. Now the car sounds like a racing car, and there's a P0219 code on the dash. Has the engine been damaged?

My AI mechanic explained the over-speed condition, politely pushed back when told the friend "knows loads about cars," and suggested a new exhaust as a safer way to get the sound. That's the kind of answer you'd want from a good mechanic.

The idea

When a mechanic plugs a laptop into your car, they're reading fault codes from the OBD port. You can do the same with a £20 to £30 Bluetooth scanner and your phone. The missing piece is someone to explain what the code means for your car, and what to do next.

I built that someone in under two days.

How it works

A real knowledge base. I collected a large dataset of OBD fault codes, each with a detailed technical explanation, likely causes and fixes.

Two ways in. Start with a code and the app does a direct lookup. Start with a description ("I'm unable to cool my car") and it uses embeddings to search by meaning. It finds the air-con entries even though the word "cool" appears nowhere in them.

A carefully layered prompt. Every request combines four parts: who the assistant is (MEC, a professional mechanic who explains things simply), the technical context it found, the conversation so far, and the latest question. It's told to use the technical data without reciting it, give a plain-English summary and practical next steps, and stay friendly when the chat wanders off-topic.

It copes with vague descriptions, several issues at once, follow-ups, and the odd question about cleaning spilt coffee off the seats.

The bigger picture

This is retrieval-augmented generation (RAG) with a hardware data source, and the shape transfers to almost any industry with fault codes and frustrated users:

• Appliance and boiler manufacturers: error codes explained to homeowners, with the right part identified before an engineer visits
• Industrial equipment: machine alarms translated into a fix for the operator on the floor
• Fleet operators: telematics data triaged before a vehicle goes off the road
• Support teams: first-line diagnosis that hands a complete summary to a human

I build these end to end: data collection and cleaning, semantic search, prompt design, the chat interface and, where needed, the hardware that reads the data. Tell me about your diagnostics problem.

More on this project

• The demo on YouTube
• The full write-up on Medium