Adam generates CAD parts from a prompt. An engineer needs to know how close the part is to what they asked for before they print or machine it. This scores one generated part on % dimensional error, then uses that score to route every request to the best model.
Built for Adam · the short version
An engineer can't print or machine a part without knowing how close it is to what they asked for.
In this sample bracket, thickness is 20% off and hole size 9% off, flagged before anyone prints it.
The data is made up. To make it real, I'd need prompts paired with the generated CAD, plus engineer edits or rejections.
Who built it
Anirudh Devella, data scientist. Before this, in production:
“I’d hire him again.” Alex Dixon, Co-founder, Bluelight (YC), managed Anirudh directly. Read it on LinkedIn
The eval layer reads the requested dimensions from the prompt, measures the generated model, and gives a % error for each one, plus a manufacturability check. Click a row to find that dimension on the drawing.
One number engineers can trust before they print. The same number is the score Adam's own CAD models can train against.
The same score, averaged over thousands of parts, ranks every model by part type. Each request goes to the most accurate model that fits the cost. If a model disappears or gets worse, the router moves on and quality holds.
Once a month, all feedback is grouped by theme and ranked by mentions weighted by how many come from paying users. The top three go straight onto the roadmap.
Next step