AI for engineering · 5 min read
A model that only makes decisions, tested on broken CAD parts
Jev does not write sentences. It answers typed questions in about a second for a fraction of a cent. So I broke 184 parts and asked it what to do with each one.
What Jev is
Jev is a model from TypeSafe AI, released in September 2026, that returns typed values instead of text. You hand it a state and a set of questions, and it answers all of them at once with calibrated probabilities.
There are only three kinds of question. A yes or no, a pick from a list, and a position on a scale.
How it differs from Claude
Claude writes. Give it a broken mesh and it will explain the problem, suggest a fix and write the code to do it, in prose you read.
Jev decides. It cannot write a word, see an image, or do arithmetic. It picks one of the options you defined, attaches a probability, and your code branches on it.
The gap that matters is cost. Jev bills $0.042 per million input tokens and answers in well under a second, so you can afford to ask it about every element in a mesh. You cannot afford to ask Claude that.
Using it on engineering geometry
The division of labour writes itself. Real geometry code does every operation, and the model only decides which operation to run. So I built a harness that breaks clean parts in controlled ways, describes each defect as a handful of numbers, and asks Jev what to do about it.
Two families of defect, which turn out to behave very differently. Openings: a hole torn in a surface, against a pipe end or a drilled hole that is meant to be there. Interpenetration: two solids sharing space with no boolean ever run, which leaves a watertight, manifold, completely wrong part.
What the model actually sees
No geometry. Twelve numbers and a paragraph of instructions, about 630 tokens. Boundary edge count, rim planarity and convexity, surrounding curvature, how many open loops the part has, how many bodies are in the file, how much volume the two bodies share, how deep the deepest penetration is. Every field is always present, so a hole reports no overlap and two overlapping bodies report no rim. If the key set changed between defect types it would give the answer away.
The harness
Clean parts are built with trimesh and manifold booleans, then broken on purpose: patches deleted at chosen curvature, rims snapped onto true circles to make convincing drilled holes, bodies driven into each other at swept depths. Eighteen parts are open by design, a pipe, a cup, a duct, panels, so that leaving a hole alone is often the right answer. Every call is cached on a hash of the request, so reruns cost nothing. The whole run is 184 calls, about 780 milliseconds each from Bengaluru, and six tenths of a cent.
Try it on nine parts
Pick a part. You see the geometry, the exact numbers the model was handed, what it decided, and what was actually wrong. Drag the part to turn it.
The part
What the model was given
No geometry. Only these numbers.
It is excellent where the numbers separate cleanly. Overlap volume and penetration depth tell interpenetration apart from a graze, and it scores 88.6% on those. It collapses where they do not. The only real signal between a tear and a drilled hole is a rim convexity of 0.78 against 1.0, and rather than discriminate it picks a default and fills everything in, including every pipe end and panel edge.
The bias flips with the question
An earlier run asked the same thing as a yes or no, is this damage. It then said "not damage" to almost everything, and missed two thirds of real tears. Reframed as a choice between four actions, on the same parts with the same numbers, it swung the other way and now wants to repair almost everything. Same model, same features, opposite failure. Whatever list of options you hand it shapes the answer it defaults to, which is worth knowing before you trust one.
How it compares to just writing rules
On the openings half, a decision tree two levels deep, fitted on labelled examples and scored on a part it had never seen, reached 80.5%. Jev reached 57.1% on the same openings with no examples at all. That is the honest comparison: if you have labelled data, write the rules. Jev's case is that it needs none, and that it arrives at an answer in under a second on geometry it has never been shown.
What I would actually use it for
Not as an oracle. As a cheap first pass in front of expensive work, where being wrong is recoverable and being fast matters.
Two-body interference is exactly that shape. Run it over every pair in an assembly, let it sort real interpenetration from tolerance noise, and send a person only the handful it flags. Ask it to judge whether a hole in a shell is meant to be there and it will confidently seal your pipe.