We build pre-training data, mid-training data, evaluations and reinforcement learning environments from the real operational data of industrial companies, for labs teaching their models to do the work that makes and keeps everything running.
Each environment starts from a piece of this work that a company has already finished. It gives the model the files and the context the people involved had, cut at the moment they decided, and grades the attempt against what the company did and how it held up.
RL environments in detailIn design and simulation our data covers CAD models and drawings with their full revision history in SolidWorks, Siemens NX and CATIA, and the CAE work behind them, from geometry cleanup and meshing to physics setup, solving and post-processing for FEA and CFD in Ansys, Abaqus and STAR-CCM+. Around every design sits its lifecycle, the change requests and change orders with their impact analyses, redlines, effectivity and where-used in Teamcenter, Windchill and 3DEXPERIENCE.
In electrical and controls engineering it covers schematics, terminal diagrams, cable and I/O lists in EPLAN, and PLC code in the IEC 61131-3 languages in TIA Portal, Studio 5000 and CODESYS. In process and plant engineering it covers process flow diagrams, heat and material balances, P&IDs, HAZOP studies and management-of-change records in Aspen HYSYS, AVEVA E3D and COMOS, through to the punch lists that close out commissioning.
In production our data covers process plans and routings, the restructuring of an engineering bill of materials into a manufacturing one, CAM toolpaths and post-processed NC code in Mastercam, NX CAM and hyperMILL, line balancing against takt time and finite-capacity schedules. In quality it covers nonconformance reports and their dispositions, 8D investigations, design and process FMEAs, control plans, PPAP submissions, first article inspection reports and corrective actions with the checks that show whether they held.
In maintenance and service it covers work orders closed with failure codes, maintenance plans, field service reports, service bulletins and warranty claims in Maximo and ServiceMax. In supply it covers RFQ packages, supplier quotes and should-cost models, supplier audits and corrective action requests, material certificates and the shipping and customs documents that move goods across borders. And in the commercial work that decides what gets built, it covers quotes for engineered-to-order machines, tenders with their deviation lists, contracts, change orders and the post-calculations that show whether a job made its money.
What you can license
Decades of real operational data at volume, from archives that never reached the internet.
Sealed industrial tasks for measuring real progress, kept apart from training.
Real industrial tasks with automatic grading, from design and change to production, quality and service.
A bill of materials is a tree, and turning the engineering view of a product into the manufacturing one is a transformation under hard constraints. An engineering change order is a structured diff to a product definition, with a stated reason, a review and an approval, and its effects run through every assembly where the part is used.
A specification, the first revision, the reviewer’s comment and the revision that answered it form a chain with a real answer at every step, which is a process signal that finished files alone can’t give.
Often the tool itself does the checking, because a CAD model rebuilds or it fails, an assembly fits or collides, a heat and material balance closes, and the schematic, the I/O list and the PLC code agree tag for tag. Much of the rest was written down as a requirement before the work began, in a specification, a control plan or an acceptance test protocol, so an attempt can be checked line by line.
Behind all of it sits the outcome on record, the first article that passed, the corrective action that held, the work order that stayed closed and the quote that won the order. Part of every collection stays sealed, and because none of it was ever published, a high score can’t come from having seen the answer.
How industrial records become agent tasksToday’s models produce geometry that looks right and can’t be made, impact analyses that miss assemblies where the part is used, root causes the evidence doesn’t support, control code that no longer matches its own I/O list, and schedules that ignore the machine that is down.
More public text won’t close these gaps, because the material that would close them was never public.
We start from the capability you are training for, whether that is CAD editing, simulation setup, change impact, root-cause analysis or quoting, and build the collection and its evaluation around it. Everything we deliver is anonymized, licensed in writing and documented, so you know what you are training on and where it came from, and the environments run in the harness you already use.
Most labs begin with a sample. Then we agree what a useful collection looks like and build from there. Evaluation sets stay sealed, so the numbers you see in six months still mean something.
Data licensingSend us the capability you’re working on, and we’ll tell you what we can build for it.