Best design.
Fewer loops.
Ready for
mass production.

Datak helps engineering teams reach the best design in fewer loops, and make it ready to build at volume: lighter, simpler, cheaper to produce, verified by physics before anyone cuts metal. A mechanical dataset today. An AI design engineer next.

A bracket cracks, a housing overheats, a manifold loses pressure. Every failure starts another loop: diagnose, change the CAD, re-run the simulation, get approval, repeat until it passes. Each loop costs days of an engineer’s time, and most parts take two or three. Then the part reaches production and starts another set of loops: draft angles, wall thickness, tooling, cycle time. That is where the time and the cost of a design go.

AI could take those loops, but it has never seen them. The internet has finished designs. It does not have how engineers fixed them. Datak records exactly that: one failure, one fix, verified by re-running the simulation, with the manufacturing changes that followed. The dataset trains and tests design AI today, and it is what our design engineer learns from: how to reach the best design, in the fewest loops, in a form a factory can make thousands of.

The figure at right is one record. Switch between the design as it failed and the design as it passed.

[ Explore the dataset → ] [ The design engineer, upcoming → ] [ Start a pilot with your team → ]

Every record verified in open solvers · no language model grades anything

Fig_001 [ Record DK-0016 · Bracket_L2 ]

↓ Drag to rotate

After · gusset 20 × 20 × 4 added · σ max 198 MPa · pass One loop instead of three · +6 g · die-cast ready, draft 2° kept

Sec_002 · One record

One failure, one fix, proven by physics

1 · Before

Design fails

Breaches a numeric limit written in the spec: stress, temperature, pressure drop.

2 · Change

One fix, one reason

Change table with old and new values, the rationale, the options ruled out, and what it meant for production.

3 · After

Design passes

Same limit, identical simulation conditions, verified from solver files.

Every record ships: STEP CAD before and after, product spec, change log, solver-native results, field data as fields.csv and .vtu, and a verification report. [ See a full record → ]

Sec_003 · The benchmark

Can AI fix a real design failure? We measure it.

Before we trust a model with an engineer’s loop, it has to pass ours. Each record becomes a task with a hidden answer. The physics grader re-runs the simulation on the AI’s design in open solvers, CalculiX and OpenFOAM, and checks the same limit the engineer had.

Eight tasks, one headline: design repair, with simulation auditing as the hardest.

[ The eight tasks → ] [ How the grader works → ]

Fig_002 [ Status · in build ]

HarnessTask per record, hidden answer, any model

GraderRe-runs the solver, checks the same limit

BaselineOur agent: read CAD, simulate, edit, retry

BoardPublic scores, before and after visuals

Scores publish with the first public release.

Sec_004 · The model Upcoming

An AI design engineer that does the loops, then readies the part for volume

Upload a design and its requirements. The engineer finds what limits performance and what makes the part hard to produce, proposes changes, re-runs the simulation to verify them, and reports the gains against your baseline: margin, mass, tooling, cycle time, cost at volume. Your engineer signs off. You see only changes that already pass.

It arrives in two steps. First a simulation auditor that checks your simulation setup for mistakes before you cut a prototype. Then the design optimizer, inside SOLIDWORKS, Ansys and NX, which takes a passing design and makes it the best version of itself for mass production.

[ How it works → ] [ Pilot the auditor → ]

Fig_003 [ Roadmap ]

NowDataset and benchmark

NextSimulation auditor pilots

ThenDesign optimizer in your CAD

GoalBest design · fewest loops · built at volume

AlwaysVerified by physics, signed by an engineer

Sec_005 · Who it is for