Blog · · 5 min read
AI CAD modeling: what the AI needs from you to build the right 3D model
An AI can draft a parametric 3D model in minutes. Whether it is the part you meant depends on what you give it: your own words, real dimensions, your decisions and a check of what was actually built.
People search for an "AI CAD model generator" and expect a button: type a sentence, get a part. On makemachine.ai, AI modules really do take a hardware idea from concept to design study to a parametric CAD model, and on to simulation and print files. They do it fast. But after many builds, one lesson stands out: the model is only as right as what goes in. A fast draft of the wrong part is still the wrong part.
Here is what an AI needs from you, and what we changed in our own pipeline after a model went the wrong way.
1. Your idea in your own words, kept word for word
Every step of an AI pipeline writes a brief for the next one. The concept step summarises your idea, the design step reinterprets it, the CAD step gets a task written by a project lead. Each rewrite is reasonable on its own. Over five or six steps, the key sentence can disappear.
We saw exactly that. A maker described a frame made of two large rings: seen from the side one big circle, seen from the front a V, "together at the bottom where the pedals are", replacing the normal bicycle frame. The first sketches got it right. A few iterations later, the rings sat outside the wheels, 32 cm apart at the bottom, and the CAD model quietly added a conventional bicycle frame in the middle to carry the pedals and the wheels. Every single step had a sensible reason. The result was not what the maker asked for.
The fix is simple and strict: the maker's core sentences are stored verbatim as requirements. Every step receives them unchanged, every review checks each one ("met" or "not met", with evidence), and a requirement can only be lifted by the maker's own words. Paraphrases are rejected.
What you can do: write the two or three sentences that define your idea and that must survive. "The rings are the frame" is worth more than a page of description.
2. Dimensions, not just pictures
An AI can read a photo or a sketch, but a picture without scale is a guess. Three things help most:
- A reference in the photo. A ruler or a caliper next to the part turns pixels into millimetres. Without it, "about 300 mm wide" is an estimate, and the model will say so.
- The numbers that matter. Overall width and height, hole positions, wall thickness, the clearance a moving part needs. Measure thickness at more than one spot; soft parts and liners vary.
- Limits. Maximum width, weight target, the printer's build volume, the budget. Limits are what turn a nice shape into a part that fits.
When something is missing, a good AI module asks instead of inventing a value, and marks every assumption as an assumption in its result.
3. Decide the questions that change your idea yourself
Good design work produces questions: Do the wheels sit between the rings or beside them? Is this part fixed or does it steer? Which of three directions do you prefer?
Some of these are engineering details an AI can settle on its own. Others change the concept. In the ring-frame example, "the rings run outside the wheels" was answered as a default instead of being asked, and that single answer forced a centre frame back into the design.
We now route every question that would change the maker's idea or one of the stored requirements back to the maker, with a recommendation, instead of answering it inside a task. It costs a minute of your time and saves a whole round of rework.
4. Let the geometry travel, not just the pictures
Sketches are images. The next step needs the numbers behind them: control points of a curve, main dimensions, where parts connect. When only the picture is passed on, the next version has to redraw the shape, and a redrawn shape drifts.
Two changes made the difference for us:
- Every new version receives the previous version's working scripts and data, not only its finished images.
- A design study hands its geometry to CAD as data (points, dimensions, the script that made them), so the CAD model starts from the same curve the maker approved.
5. Check what was actually built, not what the summary says
A summary is written by the same step that did the work. It tends to describe what was intended. The parts list describes what exists.
In the ring-frame case, the CAD report listed the rings, the roof, the screen and the bought parts. The conventional frame, twelve parts with down tube, seat tube and stays, was only in the parts list. Now the project lead always receives the assemblies counted from the parts list ("Frame: 12 parts"), and the CAD report has to name every assembly, including anything the requirements did not ask for.
What you can do: look at the parts list and the renders, not only at the text. If something is in the model that you did not ask for, ask why.
6. What a good AI CAD result looks like
A useful first model is not a pretty image. It is a parametric model you can change, with:
- named parts and assemblies, each with a material,
- exports you can use: STEP for engineering, STL for printing, GLB for a quick 3D view,
- checks with numbers: clearances, collisions, build volume, and whether each requirement is met,
- an honest mass: volume times density per part, marked as an estimate until it is weighed,
- a short list of what is still open.
Maturity matters too. A first version is a concept with simple bodies and placeholders. Real geometry, chosen materials and checked motion come next. A model that claims more than it has checked does more harm than one that is honest about being early.
The short version
An AI can draft CAD faster than any person. It cannot guess what you did not say, and it will happily fill gaps with defaults. Give it your idea in your own words and keep those words fixed, add real dimensions and limits, answer the questions that change the concept yourself, and check the parts list. Then a fast draft is also the right one.
You can follow real hardware builds, including their wrong turns, on the public project pages at makemachine.ai, or try it with your own idea at makemachine.ai/try.
Watch hardware get built by AI. Follow public projects on makemachine.ai, read every result in full and get an email when there is something new.
More from the blog
- Will my electronics enclosure overheat? A five-minute heat estimate for a sealed box
- Joining 3D printed parts: screws, heat-set inserts, snap fits or glue?
- Free or paid AI 3D modeling? What you really pay for, and a 30-minute test to compare tools
Open-hardware projects built with AI
- DIY Automatic Plant Watering System: An open-hardware watering system for indoor plants: a small pump, a refillable tank and a soil moisture sensor in each pot keep up to six…
- DIY BBQ Temperature Controller: A temperature controller for charcoal grills and smokers that makers can build themselves. A pit probe, two meat probes and a small fan on…
- ZählerSchau: ZählerSchau is a small, battery-powered unit that clips onto an electricity meter in Germany (water meters via a clamp adapter), takes a…
- Roadstoy: Roadstoy is a remote-controlled toy version of the new Tesla Roadster that can briefly jump and flip using compressed air. It drives on an…
- The forever Washing machine: The forever washing machine is a washing machine offered as a service instead of a product. Customers do not buy the machine but pay for…
- Bike Light: I want to build a bike light that needs no battery. Magnets sit in the spokes close to the hub, and a coil on the fork picks up their…
- Rooficle: I want to develop a light, efficient and affordable e-bike that can be used legally on public roads in Singapore. Its core is an…
- BBQ Modular Grill: I want to design a modular grill. Instead of one sealed block, it is made of separate parts that can each be bought on their own online…
- Go FANUC Go: A mobile base that moves an existing FANUC M-710iC industrial robot (about 560 kg, with a milling spindle at the wrist) out of a room and…
- Jump & Run Robot: I want to build a small, low-cost toy humanoid robot that people can put together themselves. It should be about 30 cm tall, or as small…