A mechanic does four things: listens, answers, remembers your car, and shows up anywhere. So does this.
Hold the phone near the noise. A listening model trained on real fault recordings names it: rod knock, wheel bearing, belt squeal, twenty nine sounds across nine systems. Then the chat picks it up from there and tells you what the fix usually involves.
On our test bench of held out recordings it named 26 of 26 correctly, every one at high confidence. When a recording is too rough to call, it says so instead of guessing. Both numbers matter.
Oil capacity. Torque for a drain plug. Tire pressure on the placard. Why the light is on. Ask in plain words and the answer comes back checked against a database of 38,729 entries covering specs, procedures, known problems, and maintenance.
The model never answers from thin air. It searches first, reads what it found, then answers. If the database disagrees with itself, it tells you that too.
| Ask about | It knows |
|---|---|
| Fluids and capacities | oil, coolant, trans, diffs, brake |
| Torque values | drain plugs, lug nuts, and more |
| Known problems | by year, make, model, engine |
| Trouble codes | what set it, how urgent |
| Maintenance | what is due at your mileage |
| Repair reality | what jobs cost and involve |
Tell it once what you drive. Every answer after that is filtered for your year, make, model, and engine. Ask about oil and it quotes your spec, not a generic one. If the database only has data for other years of your model, it says which years it is quoting. No silent guessing.
The model and the database ship inside the app. Parking garage, mountain road, the shoulder of the highway with the hood up. It all works the same, because nothing needs a network.
That is also the privacy story. There is no server. Your questions and recordings cannot leak from a place they never go.