It started with a question: what if your phone could hear what a mechanic hears? To get there we needed real cars with real problems. So we went and recorded them.
We showed up to car meets and asked: can I record your engine? People said yes. Enthusiasts wanted their turbo spool in the dataset. Owners with known problems wanted their rod knock or worn bearing to help the next person hear theirs coming. Idles, cold starts, pulls, brake stops, clunks over speed bumps. Hundreds of recordings, gathered in person, on purpose, with permission.
Then we widened the net: curated recordings of documented failures, screened by machine so narration and wind stay out of the training data. Every clip that teaches the tuner is a clip we can point to.
The chat answers from 38,729 entries: specifications, fluid capacities, torque values, maintenance schedules, known problems, and trouble codes, organized by year, make, model, and engine. It grows in reviewed batches. When two entries disagree, we hunt down the right number and fix every stale copy, not just add another opinion.
Experienced mechanics and automotive educators contributed diagnostic knowledge along the way: the pattern recognition that usually takes decades on a shop floor. That thinking is in the answers.
Nothing ships on vibes. Before an update goes out it faces the bench:
Fig 7. The bench runs before every release. A worse score means no release.
Everything runs on your phone. Not as a feature. As the architecture. The model, the database, the search engine, all of it lives on the device. We run no inference servers and keep no analytics dashboard. We cannot read your conversations because we never receive them.
Hold your phone near the engine. Press record. Get an honest answer. That was day one and it is still the whole plan.