AI that outputs parts, not pictures of parts

Every founding story starts with someone hitting a wall. Ours started with three people hitting the same one, from different directions, on different side projects. CAD software is exact but slow — you spend as much time babysitting sketches and feature trees as actually designing anything. The new wave of AI tools was supposed to close that gap and instead exposed a different one: ask an image model for anything and it hands you a gorgeous render; ask it for an M5 bolt with the correct thread pitch and you get a triangle soup no machinist would touch. Beautiful demos, unusable parts.

One of us had the background to treat that as an engineering problem instead of a complaint. Faraz had already founded Indepth Solution and spent his working life since writing production software and, more recently, AI systems — as a senior full-stack developer and as an AI automation engineer. He looked at the gap between “renders anything” and “machines anything” and recognized it for what it was: a data-representation problem. Meshes are the wrong shape for a manufacturable part. The fix wasn’t a better renderer — it was a model that speaks CAD’s actual language, parametric operations instead of triangles.

The other two didn’t need convincing the problem was real; they’d both hit it too. But they saw a case for it that went past a personal itch almost immediately. Rehan looked at it and asked the questions somebody has to ask before an idea becomes anything more than a demo — what this costs to run, who would actually pay for it, how you keep it alive long enough to find out. Salman looked at it and saw the other hard part: none of this matters if nobody outside the three of you ever hears about it, so somebody has to get it in front of the right people.

What started as a pitch for a final-year project outgrew the assignment fairly early. We kept building past the point anything was actually due, because every piece of the pipeline that worked made the underlying bet look better, not worse. A model that writes CadQuery instead of guessing at geometry doesn’t just look more like CAD — it is CAD: editable, checkable, and honest about what it can and can’t build. That’s the whole thesis, and it’s the reason we’re still doing this.