CADEX

What it does

CADEX is your CAD engineer.

Kick off your design with a prompt and any relevant reference photos.

CADEX ponders, then returns a list of requested dimensions, both in text, and annotated callouts on your reference photos, marking exactly where to measure.

Add more measurement and send follow up context until your design is fully specified and aligned with you vision.

CADEX then programmatically generates the CAD models, along with engineering drawings, manufacturing specifications, and IKEA style assembly and installations instructions.

For our demo, we used that pipeline to design a case for our beautiful Hack the North badges! We also designed a replacement foosball handle and a custom cup holder to solve problems we found around the venue.

Inspiration

AI has dramatically lowered the barrier to building in the digital world. Someone with an idea can describe what to create, iterate with a model, and produce an app, website, or tool without first becoming an expert in every technology underneath it. The distance between having an idea and being able to make it real is shrinking, but mostly in software...

On Friday night, when exploring the Hack the North venue, we stumbled on some small problems. We found a broken foosball handle, a missing ping-pong table clamp, and a lectern missing a cup-holder insert. These are all simple physical issues, but to someone without CAD knowledge, they demand a disproportionately difficult solution. Understanding how and what to dimension, physically design it, constrain geometry, iterate, and have it integrate physically all pose a barrier on people with crystal clear ideas. Frontier models are getting really good at CAD, but they are not being harnessed in a way that empowers people to actually leverage them. So we built the harness that lets AI turn human intent and real-world dimensions into something manufacturable.

How we built it

We built CADEX as a React and TypeScript application connected to a Node server and a Python CAD pipeline. The model examines the user’s photos, identifies the measurements that matter, and generates parametric CadQuery code from the dimensions they provide using OpenAI's frontier models.

Every generated model passes through an automated geometry checker before CADEX exports STEP, STL, and G-code files. We also use Three.js for the interactive preview and extract measurements from the finished STEP model to create dimensioned engineering views.

Our project uses OpenAI’s tools in two complementary ways. The measurement step sends uploaded photos to the Responses API and uses Structured Outputs to return predictable, schema-validated data: required dimensions, measurement priorities, coordinates, and visual callouts.

Codex was an absolute powerhouse when it came to shipping this project. We used our Codex credits throughout planning, implementation, testing, and debugging, but we also integrated Codex CLI as our CAD-generation agent. CADEX launches it non-interactively inside an isolated workspace, where it receives the confirmed measurements, writes a parametric CadQuery program, and executes the tools needed to inspect its own result. A concrete example is our validation loop. In early prototypes, we had to inspect every generated model manually. Working with Codex CLI, we built a checker that validates the geometry, dimensions, and hole placement, then gives exact failures back to Codex for another attempt.

Challenges we ran into

Our biggest challenge was learning that photos are useful for understanding an object, but not reliable enough for precise physical dimensions. Perspective, reflections, and hidden features made it unsafe to let the model guess measurements, so we designed CADEX to visually guide the user through additional caliper readings.

Furthermore, even with accurate measurements, the AI-generated CAD was not always dependable: it could understand the overall structure but occasionally mirror or flip the part, place features incorrectly, or produce something that did not match our intended orientation. We had to iterate on the design and automatically reject parts that failed our checks.

We also ran into problems during fabrication. We initially generated printer files using the profile for the wrong printer, then had difficulty loading and aligning the correct filament on the machine we were actually using. These issues showed us that producing valid CAD is only part of the process; printer configuration, material handling, tolerances, and physical testing are equally important.

Accomplishments that we're proud of

We are proud that CADEX became a complete photo-to-print workflow, demonstrated by the badge case we designed, printed, and presented in our demo. To our slight surprise, it fit perfectly: every screw lined up, and the case assembled exactly as intended. Seeing it work helped validated process, from uploading photos and receiving measurement guidance to generating a model, inspecting it in 3D, and turning it into a functional physical part.

Another cool feature we are proud of is the dimensioned engineering views and IKEA-style user manual, which are generated from the finished STEP geometry itself. We thought this was a thoughtful touch that makes the result feel less like a raw engineering file and more like a complete, approachable product.

What we learned

We learned that AI is most useful for engineering when it operates inside a carefully designed harness with clear constraints. The model is strong at understanding intent and proposing geometry, while the harness gives it structured measurements, fixed tolerances, examples, and automated checks. Calipers provide the physical truth, and the validation loop gives the AI a way to identify mistakes and iterate. We realized that reliable AI-generated CAD depends not only on the model itself, but also very careful consideration into building the right system around it.

What's next for CADEX

Next, we want to reduce the amount of manual measurement required by supporting reference mats and existing design files such as STEP, KiCad, and Gerber. We also want to support more printers, materials, and tolerance profiles so CADEX can adapt its output to a wider range of fabrication workflows.

More broadly, we hope CADEX inspires builders to explore what becomes possible when AI moves beyond the screen and begins working with the physical world. We want people to feel excited about a future where designing, repairing, and creating custom objects is accessible to anyone with an idea and a problem worth solving.