What a $0 AI Build Actually Looks Like

The workflow

BearingBrain answers queries like "which bearings survive 12,000 hours at 12 kN radial, 1,800 RPM, 95 °C?" For each query it filters the catalog down to up to 100,000 candidates and runs an ISO 281 rating-life calculation on every one. That loop is the hot path — the calculation that decides whether the product feels fast or slow.

We fed that exact calculation to an agentic GPU-kernel generator. The winning fused kernel runs the 100,000-candidate sweep in 0.013 ms on an H100 and 0.046 ms on an RTX A3000.

The verification

A generated kernel is only useful if it returns the same answers as the reference implementation. We verified the kernel on hardware that had nothing to do with the generation: an RTX A3000 in a workstation, against the production TypeScript reference.

• 34.8x faster than the eager reference at 100k candidates, 67x at 1M.
• Max relative error 2.8e-7 against the reference — inside tolerance.
• Latency stays near 50 microseconds as the candidate count grows tenfold.

The cost

$0.00. The kernel generator's free tier plus hardware we already own. That is the honest version of the story: no rented GPU cluster, no six-figure training run, no line-item that only makes sense in a pitch deck.

Not every build is free. But the path from "we should accelerate this" to "it is measurably faster and verified" is shorter and cheaper than the AI-services industry would like you to believe.

What it means for your project

If you have a calculation that runs too slowly, the first question is not "which model should we buy?" It is "where is the time actually going, and what is the smallest change that fixes it?" Sometimes that change is a generated kernel. Sometimes it is a better index. The field report shows the discipline either way.