Building smart geometry models @spectral_hq

San Francisco, CA
Joined September 2021
For generative AI to become useful for engineering design, we need to build new models that natively understand spatial relationships and physical reality. Excited to share our new model SGS-1, a huge step forward in this domain!
1/ We are excited to announce SGS-1, a SOTA foundation model for physical engineering design. SGS-1 enables the creation of manufacturable CAD geometry for real engineering workflows. This example shows SGS-1 in Fusion360 CAD software creating a bracket for a roller assembly.
3
11
1,283
The CSWP is a multi-hour and multi-task evaluation of CAD capabilities. Excited to share our results publicly, along with some of our research.
We built the first AI system to achieve the industry standard CAD certification for mechanical engineers. Our agent autonomously built parts and assembled mechanisms, passing the SOLIDWORKS Professional exam in less than 3 hours. Details below, including a concerning discovery.
9
473
While having our agent pass the Solidworks certification exam, we discovered some issues
Replying to @spectral_hq
Unfortunately, while preparing this announcement we identified a glaring issue: the exam exposes the answer key. The below video shows the Associate-level exam being passed in less than 30 seconds. We reported this issue to @SOLIDWORKS in advance of this post.
4
279
Many people working on AI for physical engineering haven't spent enough time looking at 3D shape modeling methods. I think there are clearly things we can learn from multi-cube.github.io/ and other works
116
The advent of NURBS was the downfall of humanity
1
1
5
781
1/ Really cool new paper on CAD sequence specific tokenization. Hard to do CAD command sequence prediction with LLM tokenizers that fragment commands.
1
7
288
Quick introduction to existing generative CAD methods and how SGS-1 compares.
1/ Current generative CAD/3D architectures offer tradeoffs between complexity and editability. We wrote a post outlining the current landscape, and where SGS-1 fits in.
9
361
Very cool results for neural CFD prediction on DrivAerNet++. Work was done by Neural Concept. Their full post: neuralconcept.com/post/from-…
1
5
220
If this is what we think it is, incredible news. GDM does amazing work.
We’re announcing a major advance in the study of fluid dynamics with AI 💧 in a joint paper with researchers from @BrownUniversity, @nyuniversity and @Stanford.
3
253
Amazing and useful application of neural methods towards solving real design problems. One of the coolest companies I’ve seen
Replying to @aphysicist
our foundation is a new kind of software that automates tool design. our AI explores the entire design space, generating options along a cost-performance pareto front that allows us to collapse months of engineering and NRE into days.
4
10
1
215
15,028
Why the hell can’t I buy ATS software without booking a demo? This seems like something that can easily be self serve
1
5
156
SGS-2 is on the way. Here's a sneak peek.
3
6
34
2,392
Rahul Iyer retweeted
1/ Excited to share that @spectral_hq has launched SGS-1, a groundbreaking generative model for parametric CAD. By creating realistic engineering designs, SGS-1 learns to model physical reality and its laws. Vibe cad’ing is around the corner.
1/ We are excited to announce SGS-1, a SOTA foundation model for physical engineering design. SGS-1 enables the creation of manufacturable CAD geometry for real engineering workflows. This example shows SGS-1 in Fusion360 CAD software creating a bracket for a roller assembly.
2
2
10
1,478
90s esthetics were wild - it was like “yayy we have NURBS now in CAD let’s make everything look like it was made by the aliens for the aliens”
90's will be an archaeologically important era in terms of massive WTFs it will cause in the future.
99
893
65
11,317
368,493
Skimmed through Figma S-1, some parts I found interesting: - Invested $200k into Bolt (StackBlitz) in April 2022 - $55m Bitcoin ETF investment in March 2024 (up $15m in a year) - another investment of $30m in May 2025 - April 2025 asset purchase (Modyfi?) of $14m - April 2025 acquisition (Payload CMS?) of $35.5m - $25.1m in compensation over 4 years - Commitment to purchase minimum of $545m in cloud hosting services over next five years
1
8
193
What would happen if machines could autonomously create in the physical world?
1
4
15
1,658
Coming soon: a new generative system that understands structured geometry
3
1
14
1,534
5
17
1,131
New structures forming. More soon.
4
1
16
1,213