Launching 26 October 2026: reverse engineering, rebuilt

What we’re launching

On 26 October 2026 we launch Oscar, OriginShape’s AI copilot for reverse engineering. Oscar reads a 3D scan the way an engineer does and rebuilds the part as parametric CAD: a feature tree with named dimensions that you can open, edit, and send to manufacturing.

• Runs: in your browser, with nothing to install.
• Takes: meshes in STL, PLY, OBJ, MA and DAE.
• Gives you: STEP, IGES and native CAD.
• Connects to: FreeCAD today, where Oscar builds the same feature tree. Connections to Fusion, SolidWorks and Onshape are coming.
• Runs on: our own AI models, Origin One and Origin One Super. Oscar is patent pending.

This post is about why we built it: what’s slow about reverse engineering today, and the ideas we think change it.

Why reverse engineering is stuck

Scanning got fast. Laser, structured-light and CT scanners capture a part’s surface quickly and in fine detail. What happens next has barely changed: an engineer opens the mesh beside a CAD tool and redraws the part by hand, sketch by sketch, measuring as they go.

The shortcuts don’t remove that work. They move it:

• Mesh-to-solid conversion wraps the scanned triangles in faces. You get a solid, but it copies the wear and noise, and there are no features or dimensions to edit.
• Automatic surfacing fits smoother faces over the scan. It looks better, but it still records the shape, not the decisions behind it.

So the bottleneck isn’t capture any more. It’s interpretation: deciding what the part was meant to be, and building it so the next engineer can change it. (For the full picture, see Scan to CAD, explained.)

Seven ideas behind Oscar

1. Rebuild the intent, not the surface

A scan measures a part as it is today: worn, and somewhere inside its manufacturing tolerance. The design behind it used deliberate, nominal sizes. Oscar rebuilds those. On the example shaft used across this site, the collar comes back as Ø36.00 mm, established from a cross-section rather than averaged from a worn surface. Why that matters is in Design intent: why a worn part isn’t the drawing.

2. The feature tree is the deliverable

A solid answers “what shape is it?”. A feature tree answers “how was it made, and how do I change it?”. Oscar’s output is a sequence of features, such as sketches, revolves, pads and holes, driven by named parameters. On the example shaft, Ø36.00 mm is the parameter d2: change it and the model updates. You can also read the rebuild as CadQuery code, which is CAD you can open, read and change.

3. Uncertainty is an output, not a failure

Some dimensions can’t be read off a scan: a worn seat, a gap in the data, a face the fixture hid. Averaging hides the problem. Oscar flags it, attaches the evidence, and waits. On the example shaft, the shoulder position at z −35.00 mm is proposed, not committed, until an engineer confirms it.

4. Agents that work like engineers

Oscar is a team of AI agents. They rebuild a part the way an engineer does, one decision at a time, in four steps: datums, design intent, features, and dimensions. Each decision is a visible step in the tree that you can inspect, change or undo.

5. Oscar proposes, the engineer decides

This is the workflow, not a slogan. Oscar suggests the design intent and the engineer approves it. The person who signs off a part stays in charge of it, and the model shows them exactly what they are signing off.

6. Judged on engineering criteria

An AI model’s confidence isn’t a measurement. We check Oscar’s output against a deterministic baseline, and we judge accuracy on engineering criteria: dimensions, relationships, and fit, not on how sure a model sounds.

7. CAD that lands where you work

A rebuilt part is only useful in the tool you design in. Oscar runs in your browser, and when you connect it to FreeCAD, it builds the same feature tree there, natively. Connections to Fusion, SolidWorks and Onshape are coming. For any other tool, there is STEP and IGES (see STEP or IGES?).

Reverse engineering, before and after

Where reverse engineering goes next

These are directions we’re working towards, not features you can use on 26 October 2026. They’re where we think reverse engineering goes once the interpretation step stops being manual:

• Spare parts without drawings. Legacy and obsolete parts become editable CAD again, ready for manufacturing, instead of one-off copies.
• Benchmarking by features, not meshes. Compare a part with another feature by feature: wall thickness, hole patterns, draft, instead of overlaying two point clouds.
• One model to rebuild and to inspect. The parametric model that rebuilds a part can also check the next one off the line.
• Every dimension with its reason. A record of why each size is what it is, so the next engineer doesn’t start from zero.
• Assemblies, not just parts. Rebuild mating parts together, so each fit is set from both sides.

Have a part with no CAD, or a scan nobody has time to redraw? Tell us about it. Follow the launch on X, YouTube and GitHub.

Further reading

• Scan to CAD, explained. Meshes, surfaces and features, and why converting a scan isn’t rebuilding it.
• Design intent: why a worn part isn’t the drawing. How nominal dimensions are recovered from a worn part, and when to flag them.
• STEP or IGES? Which CAD file format to ask for. What each exchange format carries, and why neither keeps the feature tree.

Questions