Indie AI diagnoses car engines by listening to TikTok and YouTube clips
Forget expensive mechanic diagnostics. A brilliant developer just built an open-source tool that listens to your engine's wheezing and pinpoints the trouble, trained entirely on noisy social media uploads. This is wild.
Adam Sohn launched cardiag, an open-source tool designed to analyze audio recordings of car engines to identify potential breakdowns. Instead of renting expensive diagnostic scanners, users can simply record their car's suspicious rattling on a smartphone and upload the file to a web app.
To train this digital mechanic, the developer bypassed boring corporate databases and went straight to the true home of automotive wisdom: raw videos from YouTube and TikTok where users record their cars dying in real time. A custom cleaning pipeline strips away human voices, terrible background music, and wind noise, leaving only the pure, unadulterated cry of a failing transmission.
The heart of the system relies on a frozen CLAP audio model, meaning there is no massive, planet-burning neural network training involved here. The custom trained classification layer weighs a mere 100 kilobytes, which is basically smaller than a single low-res picture of the car it is trying to save.
While it is not ready to replace your local mechanic who charges $100 just to look at the dashboard, the system boasts an AUROC score of 0.79 on noisy real-world data. When tested on clean, professional engine recordings, that accuracy climbs to an impressive 0.93, proving that TikTok's chaotic audio is the only real bottleneck here.
Even when the AI gets confused, it refuses to lie to look smart, opting to return a blunt "UNCERTAIN" response rather than confidently guessing. The system successfully places the correct broken part into its top three suggestions in 76% of cases.
An open-source project built on social media scraps is outperforming corporate diagnostic software while being completely honest about its own limitations. The era of mechanics charging a kidney just to read an error code might actually be under threat from teenager uploads.
Source: GitHub
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.