B-REpresentation · Mechanical engineering · ai supported CAD modeling
CAD was built to describe geometry, not to know what manufacturing would ask of it.
A CAD model is a precise description of geometry and almost nothing else. Wall thicknesses, hole spacing, overhang angles, tolerance stacks, whether an adjacent part will collide in assembly — none of that is carried by the model itself. It lives in the designer's head, in a standards document, or in a review meeting three weeks later.
That gap is where the cost is. Building the geometry is slow but tractable: feature by feature, constraint by constraint, a designer who knows the part can get there. What's genuinely expensive is the iteration that follows — the redesign after a DFM check fails, the fit problem found during assembly, the tolerance that was correct on the drawing and wrong on the machine. Complex geometry with free-form surfaces multiplies both: more time to build, more places for a downstream constraint to be violated.
We come to this from over 20 years of doing the work for the major vehicle manufacturer in Germany. Every part we generate has to survive the same checks our own designers apply — which is why we treat manufacturability, semantics, and geometry as one problem rather than three stages.
What we’re actually building
Four decisions that shape the system:
Function-driven input. Real B-rep geometry. Manufacturing constraints at generation time. constant interaction between the AI-driven generation and the human.
Each one is a place where the model gains information it would otherwise lose.
Function first, shape second
The input isn't a description of a shape — it's a description of what the part has to do. Mounting position, load path, fastening concept, the space it has to occupy without colliding. The geometry is determined by the function and other requirements, such as the manufacturing process, and is not specified directly. And if the function determines the shape of the component—as is the case, for example, with the visible parts of the car body—then the system treats it that way as well.
Creating 3D-Geometry with B-Representation
"We choose B-Rep — not because it is easy, but because it is hard."
Because making an AI generate not just geometry, but valid topology at the same time, is the hardest unsolved problem in computational CAD. And because only when you solve the hard problem do you earn what no other representation can give you: geometry that a simulation trusts, a machine can cut, and a technical drawing can be derived from — without a single conversion step.
Manufacturing and other constraints applied during generation
Every model is validated against real production limits — minimum wall thickness, hole spacing, overhang angles, tolerances and fits — before it reaches a person for review. Design risk that normally surfaces late gets caught while the model is still being built. The constraints serve not only to generate geometries that are valid in terms of manufacturing and functionality, but also to increase the determinism of the process.
In-time interaction between the creator/user and the intelligent behind Menta
This interaction is crucial: the user has the idea and knows what problem they want to solve with it. In addition to its primary generative function, the AI possesses knowledge of the requirements and standards for this specific production, does not overlook potential conflicts, and can simulate the required function in the background to either support or refute it. Users can focus entirely on the product and the idea.
Where we are today
Built
It comprises three built components, all of which are currently limited to a single manufacturing domain (sheet metal bent parts) across every pipeline. the first, a browser-based workspace with a three-level feature tree that decomposes a part into functional features, each of which is anchored by its own local coordinate system. This provides engineers and models with a shared, granular interaction surface. Secondly, our model training architecture is supported by a formal token vocabulary and a deterministic pipeline that derives a complete functional description directly from a STEP file. The third is a reconstruction pipeline that deterministically transforms a two-dimensional isometric view of a part into B-Rep geometry. The most important rule is that the AI makes decisions about construction, not geometry. The geometry of the system is determined by its deterministic nature, which is a fact that is not open to discussion.
Building now
In model training, we are evaluating competing methods for representing topology with the goal of making the spatial relationships within a part legible to a transformer, and we are examining language models with a visual backbone for stronger spatial understanding. In the image-to-B-Rep path, the current bottleneck is mesh generation from two-dimensional isometric views. We are pursuing a number of approaches to enhance the source sketch prior to reconstruction. The rationale behind this is that a better-conditioned image communicates proportion more clearly and should yield cleaner meshes with fewer pseudo-forms and less noise. The quality of the model's decisions is dependent on the volume of data available, and we are therefore building a system that generates synthetic training parts, describes them automatically through our existing description and topology pipeline, and feeds them back into retraining as a closed loop.
Beyond
It is worth considering how mechanical design actually proceeds today. A concept is drawn, modelled in CAD, evaluated by FEM, corrected in CAD because the simulation revealed where it fails, checked for manufacturing feasibility, corrected again because the earlier design could not be produced, re-simulated because the geometry changed — and around it goes. Every evaluation happens after the geometry exists, and every correction invalidates the evaluation before it. The loop is not slow because engineers are slow; it is slow because judgement and creation are separated in time. If manufacturing knowledge sits at the decision level and deterministic simulation runs alongside generation rather than after it, that separation collapses and the loop shortens. There is a second consequence worth sitting with. Today the design space is bounded not by physics but by engineering hours — one pursues the first workable concept, rarely the best available one, because evaluating alternatives is expensive. When evaluation becomes part of generation, that cost changes character. We leave to the reader what becomes buildable when exploring a hundred variants costs what exploring three costs now.
Why now
From static geometry to physical AI
We are entering a historic inflection point in mechanical engineering software. This is the transition from static, manual Computer-Aided Design (CAD) to native Physical AI.
The computational barrier has fallen
For many years, parametric CAD was unaffected by generative artificial intelligence because of the significant computational demands of high-dimensional Boundary Representations (B-Reps), feature trees and exact mathematical topologies. That barrier has now fallen. Exponential leaps in AI model scaling and expansive context windows now allow neural networks to process, retain, and synthesise dense industrial CAD representations at scale. Concurrently, AI has achieved a significant milestone in spatial reasoning, with models now capable of perceiving objects in three dimensions, including volume, clearances, orientation, and complex assemblies within a shared physical environment. This evolution is driven by a significant increase in research, with multimodal transformers, visual geometric backbones and dual geometry-topology graph neural networks playing a key role in converting complex B-Rep graph structures into computable, semantically rich spatial intelligence.
Native generation, not a prompt layer
While early market attempts have tried to bridge AI and CAD by wrapping legacy desktop software with basic API connectors or prompt adapters (e.g., Model Context Protocol / MCP), it has been demonstrated that this surface-level approach is fundamentally insufficient for production workflows. It is important to note that complex mechanical engineering cannot be reduced to simple text prompts. This is because such an approach would remove the precise, deterministic control and interactive feedback that engineers require for real-world manufacturing. The industry is demanding a native generative paradigm: a system designed from the ground up to support deeply interactive, co-creative design workflows. This next-generation model represents a significant advance in the field, as it natively comprehends mechanical function, thereby transcending the limitations of traditional "dead geometry". By understanding the purpose behind physical attributes, such as load-bearing ribs, heat dissipation fins, snap-fit joints and mounting bosses, the AI evolves from a mere drawing tool into a proactive engineering partner. It collaborates with human designers to co-develop complex concepts, evaluate structural trade-offs, replicate proven physical principles, and derive entirely novel geometric solutions to satisfy rigorous operational demands.
Manufacturing rules inside the generative loop
The true power of this physical AI architecture lies in its seamless integration with manufacturing rules and physical reality. By embedding domain-specific Design for Manufacturability (DFM) guidelines alongside automated real-time Finite Element Analysis (FEA) and physics simulation loops directly within the generative process, the model continuously evaluates structural stress, thermal dynamics, and material behaviour prior to finalising geometry.
Shorter path from concept to production
This closed-loop feedback mechanism creates an immediate connection between abstract vision and physical execution. Non-specialist creators who lack advanced degrees in FEA or specialised manufacturing domain knowledge can reliably generate certified, production-ready CAD data in minutes. By eliminating the friction between initial design, simulation validation and physical prototyping, this technology radically shortens hardware development cycles, democratising high-precision engineering for the global market.
Team
Valta Engineering has spent years in paid client work in automotive and industrial product development across Germany, Italy, Czech Republic and Slovakia. The company specializes in supporting major OEMs, their suppliers, and smaller companies in product development. We have extensive experience and in-depth expertise in every stage of the product development process. In the initial phase of a project, we design the concepts and create the initial CAD geometry for the prototypes. We then work on implementing the requirements and preparing the documentation for production, from series development through to industrialization. In doing so, we ensure that the developed geometry is ready for manufacturing.
Brillian is a Finnish AI company working with industrial and enterprise clients across both strategy and delivery — from deciding what is worth building to getting it into production and into daily use. Brillian brings the model engineering: the generative models behind the geometry, the data and training work they rest on, and the infrastructure that turns them from experiments into something that runs reliably.
Together, we are building what neither discipline has produced alone: deep manufacturing domain expertise fused with applied AI engineering, aimed squarely at a real industrial problem rather than a generic use case.
Track record
With MENTA-CAD, an engineer can describe a component using natural language and supplement the description with images or sketches. The engineer can then work directly with the resulting geometry, verify the result, and refine it as needed before sending it to production.
It is a browser-based CAD environment with a full-featured kernel for rendering contours. Edits affect the actual solid model rather than a visual approximation. As a result, the results can be used in downstream processes.
Multimodal input. Text instructions combined with photographs, technical drawings or hand sketches.
A deliberate architectural separation. The AI produces design decisions and a deterministic geometry engine builds the model from them. Geometric validity and manufacturability are therefore properties of how the part is constructed.
Interaction at the feature level. A change always affects only a single functional feature of the component.
Engineer in the loop. Every result is reviewed and approved by a person before it moves on. The system is built to support engineering judgement, not to replace it.
A complete in-house data pipeline for training domain-specific models, developed from the ground up for the sheet-metal domain.
This has been validated as a public demonstrator at Technology Readiness Level 7 — tested in an operational, near-production environment rather than a lab setting alone. The underlying methodology is documented in two publications, and we are currently working with patent counsel on IP protection for the core technology.
precision manufacturing
industrial automation
engineering services
CAD-AI
The ask
Everything you've seen so far was built without outside investors: funded by Valta Engineering's own resources and supported by the EU Recovery and Resilience Plan. We're now opening our first external funding round to take MENTA-CAD from validated demonstrator to market-ready product.
Capital raised will go toward three things: completing productization of the platform for commercial deployment, securing our core IP through the ongoing patent process, and building the initial go-to-market team to bring our first pilot customers onto the platform.
We're structuring this round to work for international investors regardless of jurisdiction. Get in touch for our data room, technical demonstrator, and full terms.
Get in touch
Contact
Peter Glova
Mechanical development engineer and founder of Valta Engineering (Munich, 2016); led development work on series projects for multiple automotive OEMs, where manufacturing constraints are resolved by hand, one part at a time.
Tel: +49 1743904856
Email: peter.glova@valtaengineering.de
