I built a manufacturing-aware CAD interpretation engine that uses LLMs to query any step model's geometric data, saving $10k/Yr on analysis software. This web-based CAD analysis tool is a workbench for developing bespoke DFM tools for manufacturing services.

I'm open sourcing Palmetto so anyone can access it or develop on top of it to build fast DFM apps for manufacturing services. Check it out here:

It works by abstracting any STEP file into an Attributed Adjacency Graph, turning geometry into a topology-aware graph so features, intent, and manufacturability can be recognized automatically and searched by an LLM. try it here:

We built this as the basis for our DFM tool at Digital Metal, making actual complex geometry checks for casting DFM in the browser extremely fast, without relying on outside tools or API's that are $$$. It would also work really well for CNC DFM checks and with some modifications for sheet metal as well.

the AAG of a cube is.... A cube lol

Great minds think alike!! We ditched our dependency on a SaaS product last week for routing and nesting. Our STEP analysis tool works with our die library to develop feasibility and time estimates now which is cool. It also is compatible with our python quoting scripts (iykyk) so we can finally use the same logic we use to price to later seed routing and time estimates on stuff. Our big issue is we would use a popular quoting system, and it would work wonderfully when we quoted complete cad scopes, but the sloppy non parametric line by line operations quoted from a pdf would always be a crap show to push to the floor. Now we built a way to share the same parametric quoting scripts from our MOS (erp without invoices) so we can see how well we did with the ballpark from the pdfs and test it against our CAD models while benefiting from our "if its a operation, it has a value" approach to making sure every op is costed and routed. Bend Sequence Planner with post processor support hits the our MOS next week.

Really, really cool!

Thanks Tim!

Impressionnant ! Voilà typiquement l'agilité DFM qui fait défaut chez nous. On aimerait voir ce genre de solutions émerger sans devoir affronter la lenteur administrative et les blocages français.

This is sick tech 🤯 Anyone claiming there are no startup ideas is simply not technical enough

The cost that these CAD analysis SDK providers currently charge is insane. $15K-100K/yr from CADEXSOFT, IIRC. Excited for the price reduction in this space now. (and that guy's blog) is an amazing open-source resource for this.

exactly, too expensive given the tools we have today.

I would prefer seeing a tool that automatically defines manufacturing times & cost as the design evolves (material type, process type, etc...). If this is an injection molded part, it would automatically estimate tooling cost for the mold and production cost based on tonnage and cavity count. Many other ideas, but costing and technical requirements evolution in real-time as the design matures is what is needed.

Nice work

Is this three.js graphics? What extension are you using for brep functionality?

yep, three.js for web display of the model. Using open cascade and custom c++ algorithms for feature detection, but because we abstract it into a query-able graph format, can do advanced recognition of features using the llm interface

Pretty cool ngl

Incredibly sick

godtier wtf

If you're mostly doing investment casting, isn't conventional DFM constraint less of an issue for part geometry?

@yacineMTB

what info would commonly be taken from this?

I am looking for a tool to use for PCB EDA routing with xAI, does anyone know of such a project?

This is amazing !! Can I DM?

ELI5

Very interesting!

How did you arrive at $10k/Yr figure?

Very cool. Just threw a whole dovetail fixture reference model at it and it ate it up lol

Connor's engine interprets CAD; LLMs query geometric data.

Trying to wrap my small ME brain around how this works🙃

Interesting

Following