# Spectral Labs Spatial intelligence for engineering physical systems - Team: Pranav Parthasarathy (co-founder), Rahul Iyer (co-founder) - Founded: 2022 - Invested: 2024 - Links: [Website](https://www.spectrallabs.ai), [LinkedIn](https://www.linkedin.com/in/pranav-parthasarathy-82674811b/) - Field: Materials and the built world ## The problem: How do you teach an AI to design a part that actually fits? ### How things get designed Almost every manufactured object starts life in CAD, computer-aided design: software for creating, changing, analysing and optimising a design. What comes out is usually a set of files that go on to printing, machining or other manufacturing steps. Serious CAD models are parametric. Dimensions are parameters you can change later, and the model updates to match. Stretch a shaft from 100 mm to 200 mm and the hub bolted to its end moves with it, while the drawings follow along. Underneath, most solids are stored as a boundary representation, or B-rep, which describes a shape by the limits of its volume: faces, edges and vertices, each tied to an exact surface, curve or point. When one program hands a model to another, a standard called STEP often carries it. Compare an STL file, the everyday format of 3D printing. It is a raw, unstructured skin of triangles with no scale, where the units are arbitrary. Fine for printing, useless if you want to change a hole's diameter. Before anything gets built, engineers often test the design with the finite element method, which chops the part into small, simple pieces called finite elements and solves for how the whole behaves. Done well, it means fewer hardware prototypes and a faster, cheaper design cycle. ### Why it is hard **Models that can't see space.** Vision-language models are good at naming and describing objects, and often bad at spatial reasoning. Some appear to fake it, leaning on relationships memorised from training data instead of real spatial understanding. Fine for captioning a photo, fatal for a bracket that has to line up with four bolt holes. **Code is not shape.** One way to get CAD from an AI is to have it write the sequence of sketch and extrude commands a designer would use. These outputs are the most useful for engineering, since they respect dimensions and can be edited. But the model has to map an image onto commands that only turn into geometry after the CAD kernel processes them, and the commands themselves are not spatial. It's like describing a sculpture by listing chisel strokes. **Pretty, but dumb.** Other models generate 3D shapes quickly and in great detail, but as meshes, not parametric models. You can 3D print them. You can't edit them in normal engineering workflows, and you can't manufacture them. Scanning a real part has the same problem: the measured points lack the topology and the design intent that made the part what it is. **Not much to learn from.** Good CAD data is scarce. One widely used dataset of CAD construction sequences holds 178,238 models, supports a limited set of operations, and is full of duplicates with little variety. A larger collection has about a million CAD models. Meshes are the most plentiful 3D data, which is exactly the format engineers can't use. ### What Spectral Labs is after Spectral Labs starts from a blunt premise: existing foundation models lack built-in spatial understanding, so they can't do much of the real work in engineering design and manufacturing. It trains large models meant to engineer physical systems directly. The field so far sits on a frontier that trades complexity against editability. Models either make intricate shapes nobody can edit or editable shapes too simple to matter. Spectral is trying to get both at once. ### How they go at it **Pictures in, parts out.** Its first model, SGS-1, takes an image or a 3D mesh and returns a B-rep part as a STEP file that can be edited in ordinary CAD software. Images are the chosen input because text is an imprecise medium, and sketches and dimensioned drawings are already how engineers talk. **Real cylinders, real planes.** Earlier B-rep generators describe every surface with B-splines. SGS-1 produces actual primitives where they belong, planes, cylinders, cones, spheres and tori, which makes the output much easier to edit directly. **Parts that fit.** Given a partial assembly and a description or image of a bracket, SGS-1 proposes a bracket for that context, which a designer imports and adjusts until it fits. It can also turn scans and STL files into parametric STEP files without human input, automating a good deal of reverse engineering. **Honest about limits.** SGS-1 struggles with organic shapes and complex curvature, with very thin structures, and it can't build full assemblies in one go. Next on the list is reinforcement learning with feedback from physical simulation. ### Still open Why do vision-language models lose track of space? One analysis found that the image tokens inside these models carry much larger magnitudes than the text tokens, drowning out the signal that encodes position. Targeted fixes predictably restored some positional sensitivity. Whether patches like that add up to real spatial reasoning, or whether it has to be learned from geometry itself, is unsettled. Can a model learn physics by checking its own work? The finite element method is the standard tool for working out stresses and deformation in solid parts, which makes it a natural judge for an AI's designs. Using physical simulation as a training signal for design models is still an early bet. ### Words used here - **CAD**: Computer-aided design: the software engineers use to model parts before they are made. - **parametric**: Built from editable values like lengths and angles, so changing one number reshapes the model. - **boundary representation**: A way of storing a solid as the exact surfaces, edges and corners that enclose it. - **STEP**: A standard file format for exchanging precise 3D product data between CAD programs. - **STL**: A simple 3D file format that stores only a surface made of triangles. - **finite element method**: A numerical technique that splits a part into many small pieces to predict how it bends, heats or breaks. - **B-splines**: Flexible mathematical curves and surfaces that can approximate almost any smooth shape. ## About Spectral Labs Spectral Labs is building AI reasoning models for engineering physical systems. The company focuses on spatial intelligence — training models that can reason over 3D geometry and physical constraints, not just understand language or pixels. The target domain is physical design: manufacturing, robotics, and the engineering workflows where better tools have real-world consequences. Spectral is building the datasets and models required to reach state-of-the-art performance in this space from the ground up. ## Sources 1. [Computer-aided design](https://en.wikipedia.org/wiki/Computer-aided_design), Wikipedia 2. [Solid modeling](https://en.wikipedia.org/wiki/Solid_modeling), Wikipedia 3. [Boundary representation](https://en.wikipedia.org/wiki/Boundary_representation), Wikipedia 4. [STL (file format)](https://en.wikipedia.org/wiki/STL_(file_format)), Wikipedia 5. [ISO 10303](https://en.wikipedia.org/wiki/ISO_10303), Wikipedia 6. [Finite element method](https://en.wikipedia.org/wiki/Finite_element_method), Wikipedia 7. [Reverse engineering](https://en.wikipedia.org/wiki/Reverse_engineering), Wikipedia 8. [Beyond Semantics: Rediscovering Spatial Awareness in Vision-Language Models](https://arxiv.org/abs/2503.17349), arXiv 9. [DeepCAD: A Deep Generative Network for Computer-Aided Design Models](https://arxiv.org/abs/2105.09492), arXiv 10. [ABC: A Big CAD Model Dataset For Geometric Deep Learning](https://arxiv.org/abs/1812.06216), arXiv 11. [Foundation Models for Physical Engineering](https://www.spectrallabs.ai), Spectral Labs 12. [Introducing SGS-1](https://www.spectrallabs.ai/research/SGS-1), Spectral Labs 13. [The AI CAD Frontier](https://www.spectrallabs.ai/research/The-AI-CAD-Frontier), Spectral Labs