Autopilot
Import a scan. That's the whole job.
Turning a raw scan into a finished part normally takes a chain of separate tools, each run and tuned by hand. Autopilot replaces the whole chain. Import a point cloud, or simply save a scan, and the part comes out cleaned, merged, meshed, inspected and exported without a single click.
Doing it by hand
7+ steps- Open the scan
- Pick noise filters and tune them
- Crop out the table it sat on
- Line up each extra scan, pair by pair
- Merge, and hope the fit is right
- Choose mesh settings and build it
- Export, then inspect it somewhere else
With Autopilot
1 step- Import the scan
Then, on its own, Autopilot
- cleans it and removes the table
- merges extra scans, but only when they genuinely fit
- builds a mesh at true size
- inspects it against CAD if you've added a model, with a pass/fail report
- exports STL, PLY, OBJ or GLB
This is no blind macro. It makes the judgement calls an expert would. Filters scale to each scan's own point spacing, so one set of settings works at any resolution. A scan is merged only when it adds new surface, lines up confidently and agrees with the others; an ambiguous fit is flagged, never guessed. Point it at a folder and every part saved there comes out finished.
At a glance
Accurate enough for an inspection report
It runs on an NVIDIA DGX Spark and serves any browser on the network. Six guided steps take a part from Scan through Clean, Align, Mesh and Measure to Export. The inspection toolkit covers true size along the part's own axes, nine tools from calipers to flatness, automatic thread identification (metric, UNC, UNF and BSP, with ISO tolerance classes), a colour-coded comparison against CAD and printable inspection reports.
One rule governs everything: no step may quietly move, stretch or thin out the data, and every step reports what it changed. That rule is why a number on its inspection report can be trusted.
The journey
From a locked-down scanner to a full workstation
-
September 2026
A precision scanner, locked to Windows and macOS
The scanner is a metrology-grade blue-laser unit. Its software runs only on Windows and macOS, and its maker publishes no developer kit. I set out to put scanning, processing and inspection on an NVIDIA DGX Spark, an ARM64 Linux machine, and reach all of it from any browser on my network.
It had to measure parts, not just make pictures of them. A scanning tool that quietly changes a part's size is worse than no tool at all.
-
Reverse engineering
Decoding the scanner's USB protocol
With no documentation to go on, I captured the USB traffic between the scanner and its own software and decoded it. The scanner turned out to be a small Linux computer in its own right: its cameras stream as standard USB video, and an undocumented command channel reads its factory calibration and switches the lasers, exposure and gain.
Then I wrote a native driver from scratch. It replays the start-up sequence, reads the calibration at every connect, streams at about 89 frames a second and turns camera images into 3D points across 12 worker processes.
What I foundThe vendor's own project files quietly store every frame's points and position. I turned them into frame-by-frame ground truth, so my driver was validated against the vendor's output, not guesswork.
-
Accuracy
From millimetres to micrometres
My first driver measured a flat board with 330 µm of error. Locating the centre of every laser stripe to a fraction of a pixel cut that to 31–40 µm.
A deeper fault hid in the stereo matching: it paired the wrong laser lines between the two cameras, throwing points off by 11–44 mm. Identifying every line from the scanner's own factory calibration eliminated that entire class of error.
-
The workstation
Six steps, one obvious action each
Around the driver I built the full application: a FastAPI engine that runs heavy jobs in worker processes, and a React and three.js front end whose 3D viewer pivots on the point under your cursor and snaps measurements to full-resolution data. Every edit creates a new model and keeps its lineage, so any result traces back to the exact scan it came from.
Then I built Autopilot on top: the same steps chained end to end, with the checks a careful inspector would make, so a scan goes in and a finished part comes out.
-
The AI agent
An AI agent that runs the whole app
A language model running on the same machine operates about 28 of the app's tools. Ask it to “measure the head height” and it draws the dimension on the model. Ask it to “clean this scan and mesh it” and it does. It highlights regions described in plain language, can drive the scanner and turntable, and reads photos of the real part to choose the right alignment. It confirms before anything irreversible, and no data ever leaves the machine.
Then I made it fast. Benchmarking and tuning models on the Spark cut a request from 162 seconds to 10.6.
-
Proving the numbers
Tracing a bolt's error to its real causes
When a bolt's measurements didn't match its spec, I didn't guess at a fix. I built a ground-truth bench of synthetic parts with exactly known sizes to pin the error on either the scanner or the software. The software held up: cleaning moved no point, merging was a pure rigid move, and the mesh tracked the scan to within a few micrometres. The real causes were subtler:
- The size readout used the scanner's axes, so a tilted bolt read 94.6 × 40.0 × 95.5 mm when it is really 107.2 mm long. Sizes are now measured along the part itself.
- The bolt's 12-sided head let two scans line up one flat off, shifting the length by 0.26 mm. Merges now detect that kind of symmetry and flag it.
- The scanner's own scan-to-scan consistency is a few hundredths of a millimetre, so new accuracy tools now separate scanner error from software error.
That investigation became a permanent safety net. About 300 automated tests now generate parts of known geometry, add noise, debris and a table, and verify that cleaning leaves under 0.1% stray points, merging recovers the pose within 0.2° and meshes stay within 0.15 mm.
What I foundSeparate the instrument's error from the software's before fixing anything. Neither the scanner nor the geometry code was at fault. The bugs lived in how size was reported and how an ambiguous merge got accepted.
What's next
Pushing accuracy further
Still in progress
I'm still building this
CloudClean is an active build, and these are the next milestones.
- Take marker-based tracking onto real stickers. On synthetic data it already tracks to under 0.15 mm.
- Track down and close a ~0.16 mm disagreement between the scanner's two laser-line families.
- Go head to head with the manufacturer's own scan of the same part.
- Prove out Bluetooth turntable control, with rotate-and-tilt programs synced to capture, on real hardware.
Expect further improvements and changes to this project as it develops.