Bringing AI to mechanical engineering | FaenoCAD (Claude CAD) | Ideal Forms CAD Services (https://nitter.gq/t.co/KYTZWDHUZe)
Portland, OR
Joined June 2011
- Tweets1.5K
- Following608
- Followers1.8K
- Likes836
Pinned Tweet
Claude CAD Fable designing Robit V0. Full interconnect is routed and packaged in CAD.
From schematic, to bench layout, to packaged layout. All from one netlist source of truth. All checked for interference and minimum bend radii. All executed in build123d code-CAD.
I'm gonna keep posting about this until it sinks in. This is a whole new world. Agentic mechanical design engineering is here!
Just finished CAD design for a client's invention using my Faeno harness.
The best test for AI CAD is designing something which has never existed before.
I know everyone's on about Opus 5.5 and GPT-6.1 whatever the name is. But good ol' Fable 5.1 understood the assignment. It was beautiful.
Docs from the client were worse than usual--drafted by chatGPT, likely the free version. Technical drawings from image generators are uniquely horrible.
Lots of AI CAD demos are impressive parts which have lots of reference information online (robots, jet engines, raspberry pi cases). Or they are very simplistic ~5 feature parts (bracket with holes, flange with holes).
New design is a different game.
Here's the problem with one-shot prompt-to-CAD in the mechanical engineering design space:
Design intent, in ME, can't solely come from the technical specification. The constraint space is much too large to fully lock down; it is almost always the case the design is under-specified in some fashion.
Take the bike crank below and ask yourself: what defines the fillet size on the arm? If a machinist comes back and says it's hard to machine, but you believe your customers really want a particular look, how do you make that call?
How about the cross-sectional size of the chainring arms? How do you make the call between stiffness, weight, and manufacturability?
This is what I'm talking about when I talk about "opinionated design". All original design in the ME space is opinionated.
One-shot agent AI is great at *unopinionated* design, where things are largely constrained by the prompt. It is excellent in highly constrained space, such as computer programming where you are largely assembling pre-built blocks from libraries.
It is very difficult to get opinionated design out of agentic AI.
So nice to just give Claude a redline of a P-ID with a summary of changes and do other stuff while the schematic is updated.
The future of software is in-line adaptation to accommodate individual challenges.
4860 holes is a lot for any CAD system. Claude adapted the FaenoCAD viewer backend in real time to switch to Delabella meshing from the default under these conditions.
The parameters in the parameter pane is straight from code. Preview in the viewer, save it to code if you like what you see. This way, the agent is charged with building features; the user can fine-tune geometry. Viewer automatically passes the changes back to the agent to close the loop.
Lots of building with prompt-sketches with this one. Claude 4.x had trouble understanding prompt-sketches showing complex revolved sections. No longer.
The fuck is going on with this astroturf campaign by @skyfallai ?
Why is every other post on my timeline some paid partnership general-interest poster with a large following spouting marketing speak as if they've actually used the product?
I can see the promise of this approach. Before the current agent models were useful for CAD (pre-Claude Code), I was exploring using an autoencoder to encode b-rep model features in latent space; have a fine-tuned LLM do operations in this latent space to create new geometry.
Reading their report, it seems they are doing something on a similar vector. Obviously pretty good results from this approach.
I am concerned about how they are marketing this. Inferring parts inside an assembly from a global picture doesn't seem to be useful. Guaranteed to be wrong with every assumption.
Also, the pause-and-zoom-in applies here. The assembly deconstructions are objectively pretty bad.
Introducing SKY-001, the first Engineering World Model.
On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested:
-> 51-125x faster
- > 150-210x cheaper
-> Higher IOU than all frontier models
SKY-001 lets you dream a physical object before you build it.
18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch).
The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.
Funny thing. A bit ago I was worried a bit about the FaenoCAD viewer not having a particular feature. Cuz, it is pretty stripped down.
But then I was thinking, if I was building something and needed a feature I didn't currently have, say, cross section on an arbitrary plane, I could just add it in, on the fly.
This is the future of software: Say I release FaenoCAD Viewer opensource with a default set of features. If a user needs a new feature, they just have Claude/Codex/Pi/whatever build it for them. They don't bug me about it, they just change their copy to their liking.
Fully custom software for everyone.
My process skid client asked for a simulation. Basically taking the P-ID and show fluid paths with valves in different positions. They asked if it were possible.
Yes! Easy! P-ID has a netlist. Once I have CAD for rev2, I can show this in CAD as well.
Guys, this is a low-key digital twin!
At my prior company, we had a whole software team dedicated to simulators such as this.
So, Claude saved Gmail's ass. Good thing. I've had this account for literally 20 years. I have one of the original gmail addresses coming out of restricted beta. Would be a shame to give it up.
Did you know you could have bog standard Claude desktop create a scheduled task which sorts your email? It also helped me set up filters by going through the last three weeks of email.
Inbox: saved
I have the scheduled task, every hour, go through email, tag emails as as "needs me", "FYI" or "Low value". Doesn't move the email or delete things.
Idea is I can use this to tune the "important" filter, which, despite Google being one of the Big Four, is incredibly stupid to the point of uselessness.
Not ready to outsource all my email reading, but I was breaking under hundreds of emails every day pouring in from desperate politicians and businesses pushing shit to buy. So many "subscriptions" I never signed up for.
Guys, I'ma gonna drop gmail. I'm gonna do it. Holy FUCK. 150 "important" emails and they are all political ads or trying to sell me something.
How the mighty Google has fallen. They are selling my inbox now!
I am in LOVE with the FaenoCAD viewer. It is super light weight. Runs local. It's got a comprehensive MCP for agent building. Builds native in build123d on OCCT8. Secondary kernels with CadQuery (OCCT7.8) and SMLib.
Of course, it can open/export STEP and STL as well.
It's as parametric as you want, with a parameters tab which takes variables directly from code. User can twiddle variables, save them, the agent knows exactly what was changed. Can even pass a message to the agent with the design intent. Don't like the parameters offered? Just ask your agent to adjust. Easy.
I'm considering open-sourcing it, but man, there is a lot of (mostly half-assed) open-sourced AI CAD out there competing for views. And I know nothing about maintaining an open-source project and the projects tend to suffer feature bloat.
Almost a year ago I started my journey with "Claude CAD". Started with using Solidworks to create a STEP and giving it to Claude Code to modify. Take the STEP export from Claude Code and opening it in Solidworks.
I then created a viewer special for agentic CAD building because straight text-to-CAD is awful. Language, by itself, is horrible for building geometry. Without a viewer, there is no way for either agent or user to holistically understand what the code is building. No way to even talk about it.
We've come a long ways since those early days. Now making real parts for real money.
The Matrix really nailed a bunch of stuff right about our current AI SOTA.
-super-agents running agent swarms aimed at a specific goal
-agents each containing different memory contexts and talking to each other in English
-even the red pill blue pill scene where in each glasses lens Neo is presented a single different pill is a characteristic image generator mistake
I’ve always wondered. We take our sense of touch for granted, especially with robotics.
Have you ever had hands so numb you had to look over at what you were grabbing for? I remember a road bike race where temperature was in the 40s and it was raining and all I had was arm warmers. My hands were numb up to my elbows. I had to look over at my shifter to shift and even with that I could barely complete the movement.
Point is, we grab things all the time without looking. We rely on our sense of touch. Not just bumping into things and getting force feedback, but feeling the fine textures of what we are touching. Ever wonder how we are able to grab a slippery plate while doing dishes without it moving? It’s because we can feel the plate slipping a micro amount and adjust grip, without vision or force feedback, before momentum builds.
Who’s working on robot control just using touch? I am super curious.
I’m sitting here, practicing guitar and not looking at hands. If my fingers were numb, no amount of muscle memory will have me playing the correct strings and frets. I wouldn’t even be able to hold the pick. Regardless of vision feedback I could apply.
Teaching a robot a new skill is a lot messier than the final demo makes it look.
Lots of failed grasps. A few unexpected mistakes and eventually, moments where it actually works.
This is what learning in the physical world looks like with MARS from @innate_bot
Super interesting. In SMLib (via smlibpy) fillets are both more robust than OCCT, and apparently somewhat non-deterministic with some new and different failure modes.
In the below compressor wheel, the script is much less convoluted in smlibpy than in build123d with regard to the root fillet (cause of all problems).
Three issues though, and they are a bit weird:
1) Fillet chains crossing the seam line of the core fail. Seam needs to be placed where it doesn't intersect any of the blade root edges
2) Fillets which cover the seam line of the core build only in G1 mode (standard rolled fillet)
3) There is a non-deterministic fillet failure on the blade opposite of the seam where it fails on 1 out of 5 builds. My script had to build-in a self-healing routine.
Verdict: Better-ish than OCCT, but odd edge cases, which in a complex design, aren't really edge cases.
I stated in a followup post that OCCT8 fixes most of the former fillet issues with this part, so maybe OCCT is a better fit. No non-determinism, no issues with seam lines. On the other hand, the SMLib fillets are much better formed, when they form.
Turns out, OCCT8 fixes all the former issues with 7.8 without introducing new ones. Only one minor workaround regarding coincident surfaces and no non-deterministic behaviors.
I'll keep watching SMlib, but for the moment, build123d 0.13.0 with OCCT8 wins on fillet behavior.
Super interesting. In SMLib (via smlibpy) fillets are both more robust than OCCT, and apparently somewhat non-deterministic with some new and different failure modes.
In the below compressor wheel, the script is much less convoluted in smlibpy than in build123d with regard to the root fillet (cause of all problems).
Three issues though, and they are a bit weird:
1) Fillet chains crossing the seam line of the core fail. Seam needs to be placed where it doesn't intersect any of the blade root edges
2) Fillets which cover the seam line of the core build only in G1 mode (standard rolled fillet)
3) There is a non-deterministic fillet failure on the blade opposite of the seam where it fails on 1 out of 5 builds. My script had to build-in a self-healing routine.
Verdict: Better-ish than OCCT, but odd edge cases, which in a complex design, aren't really edge cases.
Not bad, not bad. Smlibpy on the left, build123d on the right. Looks the same, right?
Which, of course, is the point.
This is our cad-query inspired Python wrapper for SMLIB. This one is 100% free and open source. We ship it with our MCP. It is in early beta (macos ✅ linux ✅ windows follows shortly).
→ smlibpy.classcad.ch
→ github.com/awv-informatik/sm…
