Every chip, every lithography machine, every aircraft and every weapon system is built from parts that someone learned to make by hand. Those parts come from shops of ten to two hundred and fifty people.
An estimator determines what a part requires. A programmer or work-preparation lead translates that into operations, setups, tooling and machine choices. Both build years of unique knowledge and experience, but that knowledge is rarely captured in a structured way.
ERP systems record orders, materials, hours and cost. CAD describes the design. CAM translates operations into machine programs.
What sits between those systems is far less well captured: why an experienced machinist chooses a given operation, which setup makes sense, which machine fits, and what a tolerance costs in time and money. That is the knowledge layer Subduxion models.
Three technical challenges
Data coverage
Historical cycle time alone is not enough. A model that truly has to understand production needs information about the part, the operations, the machines and the actually measured outcomes. It is the connection between that data that lets it go beyond predicting time alone.
Physics, not pattern matching
A neural network can learn correlations from historical data, but a manufacturing process is also governed by physics: geometry, forces, heat, material behaviour, machine kinematics and process limits. That is why our architecture separates geometry, process knowledge, physical models and learned corrections.
Every process is different
Milling, turning, wire and sink EDM, grinding and sheet metal do not follow the same physics. That is why we do not build one large model to solve everything. We work with one shared representation of the part and the factory, with specialised models per manufacturing process.
Four foundations
Learning from real production, part requirements linked to the right geometry, working on top of the systems a shop already uses, and confidentiality: four things that have to hold before a calculation is traceable to its inputs and assumptions. We publish each one with its source.
The four foundations
Learning from real production
A calculation predicts what an operation will require. Production shows what actually happened. Your machine already logs its own cycle time, over MTConnect and OPC UA among others. Since 12 September 2025 the EU Data Act gives you the right to that data in machine-readable form; from 12 September 2026 new equipment must offer that access by design. By putting planned and realised time, machines, materials and operations side by side, the model can learn what actually works in a specific shop. That feedback loop is part of our next development phase, not something we present as running everywhere today.
Requirements linked to the right geometry
A tolerance, fit or surface finish only means something once it is clear which face, hole or other geometric element it applies to. Blake is learning to connect information from 2D drawings to the 3D model. Working directly from a full model-based definition and PMI is a next step, not something we claim today.
Built on top of the systems a shop already uses
Blake is designed to work alongside the ERP, CAD and CAM systems a shop already uses, not to replace them. The model brings together information from those systems around a single part and the decisions needed to make it. Start from an upload of a 2D drawing or 3D model; additional intake channels are part of the rollout.
Built for sensitive manufacturing data
Technical drawings, 3D models and production data often carry confidential knowledge from both supplier and end customer. That is why we apply high security standards to storage, processing and access, and develop AI with clear boundaries around confidentiality, control and responsible use. Where needed, processing can be set up within the EU or within the customer's own environment.
Master plan
Understand how parts are actually made, from geometry and requirements to machines and the decisions in between.
Put that knowledge to work where estimating and work preparation already happen today, with Blake, and let the specialist confirm every decision.
Turn production outcomes into new knowledge: compare every prediction with what actually happened on the machine, and use that to make the next decision more accurate.
A world model that proposes how a part can be made, which operations, in what order, on which machines and under what conditions, so people prepare manufacturing work on knowledge instead of word of mouth.
How the model grows
First generation
Geometry, operation, machine and cycle time. The model learns what a geometry means for the manufacturing process, which operation belongs to it, which machines are suitable, and how much time that is expected to take.
Next layer
Material, tooling, setups, tolerances and measured production outcomes, so the prediction becomes not only part-specific but increasingly factory-specific.
Further development
Process physics and multi-step manufacturing routes: forces, heat, tool wear, vibration, deformation, failure modes and the dependencies between successive operations.
Longer term
An intelligent process planner: a system that can compare multiple possible manufacturing routes and propose an operation sequence based on manufacturability, time, cost and available production resources. People confirm; the planner never executes automatically.
Measuring how much of a part the model truly understands
A simple part says a lot through its shape alone. As tolerances, datums, material treatments and interdependent requirements increase, understanding how the part has to be made takes more and more manufacturing knowledge. That is why we are developing a benchmark that measures how much of that manufacturing intent a system can correctly reconstruct.
A plate with three through-holes, general tolerances, one material: the shape itself tells nearly everything.
A turned-milled part with two fits (for example H7/g6) and one surface-roughness requirement on one face.
A part with a simple datum reference (one or two datums), position tolerance on a hole pattern, and a separate heat treatment.
A part with a full datum reference frame (three datums), an assembly with a bill-of-materials line, or a sheet-metal part with bend-order-dependent tolerances.
A part with composite tolerancing, a simultaneous requirement across multiple features, and a virtual-condition calculation that affects the fit.
The C4-C5 zone (highlighted) is where the gap is today: a direction, not a finished result.
The method is final. The numbers follow once the public set has run and been re-scored.
The axis is fixed in advance.
The definition of complexity is set before a test runs, not after.
C comes from the input, never from our own output.
A part's complexity is set by the generator, not by what our own system makes of it.
We generate the parts ourselves.
The ground truth is fixed in advance, and no real customer part ever goes through it.
Bits and counts.
Every result shows both the information measure and the raw count behind it.
Confidence bands.
Every result shows a range, never a single point alone.
Where we're heading
Establishing the baseline across C1 through C3, every outcome with a traceable source.
Closing the gap at C4: binding requirements to the right face, assemblies with a bill of materials, and sheet metal added.
From reading to judgment: fixturing sequence and process choice for tolerance chains, raising the curve at C5.
Closing the loop with real shop-floor outcomes, so estimates become measured relationships, and widening the axis to more process families.
Every step gets the same measurement, with the same public recipe. What is not measured is not shown here.