Stories
Eight months of scanning, modeling, printing and refitting parts for Irko's car. Every one had to match a car that was already built.
When you design for a car that already exists, the car is always right and the drawing is always wrong.
Between December 2025 and July 2026 I designed and built 22 parts for the car of Irko, a multi-platinum audio engineer. A rear camera and light housing, light bars, a wireless charger, scent holders, footwell panels, badges, hub caps. None of them existed off the shelf. That is why he came to TraceHaus.
There are no drawings for the inside of a finished car. There is only the car: its curves, its clips, the gap behind a seat. A part that looks right on screen and misses on the car does not go on. So every job here starts with the car, not the screen.
Each part began with a 3D scan of the place it goes, seven scan meshes across the job. I modeled every part in SolidWorks directly against the scan surface, in the scan's own coordinates, so the part and the car share one frame. Then I printed it, took it to the car and fitted it.
Then I revised it. Not from the screen, from the fit.
The rear camera and light housing shows the loop best. It sits above the plate and carries the reversing camera and a light. The early revisions settled the camera window. V11 added the lower brackets. A B version followed in May, and that is the one on the car, after 13 revisions.
Inside the car it was the same loop. The scent holders are a matched driver and passenger pair, 28 CAD files between them. The wireless charger went through three printed prototypes before the final arm and ring. Long light bars were split into sections sized for the printer, each modeled on a scan of the panel it mounts to.
Not every print made it. The timelapse below stops just before a side print failed. That is part of the job: every print is a test of the last revision.
By July, all 22 parts were designed, made and installed, from more than 200 CAD files. The last file in each part's folder is the revision on the car.
Looking back, most of those eight months were not design. They were measuring, modeling against a scan, checking, and doing it again. That is the slow, exacting work I think AI should be doing. It is what I am building with Project Ground Truth: AI that reads the scan, builds the CAD and gets checked against the real part, so a person's time goes to the decisions only they can make.
Rebuilding two hand-laid fiberglass hoods as moldable CAD, with an AI agent doing the modeling and a log of everything it got wrong.
A part can pass every check a computer runs and still be wrong.
A restoration client needed two hand-laid fiberglass hoods for the same car remade. There were no drawings and no molds left. Hood A has a raised dome, twin scoops and rows of louvers. Hood B has heat-extractor vents and an inner panel bonded to its back. The mold shop needed geometry it could use: drafted plugs to machine, and solids with every opening in place.
I scanned both faces of both hoods: four meshes, about 12 million triangles. Before any modeling, I read what the maker meant. The creases were straight. The flank angles were constant. The louvers were meant to be even. Under all the custom work sat a stock crown.
Then the modeling went to an AI CAD agent, working in Python on an open-source kernel, with one rule: log every miss. It built each hood as one parametric script, where every crease, angle, pitch and opening is a named number. I reviewed every part and made every design call.
It missed. An inverted vent insert and a hidden void both passed the kernel's validity check. As far as the software was concerned, the parts were fine. Only cutting sections through them and laying the scan on top showed they were not.
Some numbers cannot come from a scan at all. Fitting laminate thickness from the two scanned faces gave 17.4 mm, then −0.7 mm. One caliper reading beats any amount of fitting, so thickness stays a placeholder until someone measures it, and the report says so.
The originals had errors of their own. Hood A is about 1 mm out of symmetry left to right, and its hand-made louvers wander by up to 5 mm. The CAD is symmetric and evenly spaced on purpose, so that floor is measured first and the CAD is judged against it.
Some calls were not geometry. The owner said the louvers were open; the underside scan shows them closed. The model follows the owner, keeps the choice behind a one-line toggle, and flags the conflict.
The two hoods' outer edges agree to within about 2 mm, so they share one crown: hood B's top is the stock surface, and hood A is that surface plus its raised centre. Edit it once and both follow. It costs hood A about 0.2 mm RMS, and that number is in the report too.
In about a week the client had seven CAD solids, parametric scripts and a deviation report that tags every dimension with where it came from. Hood B's skin sits at 0.58 mm RMS against the scan outside the vents. Hood A is 1.26 mm RMS, against an original that is itself about 1 mm out.
The agent logged 23 failure modes. Nine are fixed or worked around. Ten are still open, and they are published anyway, along with the fit check that only happens after the first pull from the mold.
Later, both hoods went back onto a scan of the same car's engine bay. The car became the shared surface the two hood scans lacked, and it started answering the thickness question: hood B's inner panel sits about 6 mm low, which points at a thinner laminate than assumed.
This is the split I want. The model does the slow, exacting work: fitting, sectioning, rebuilding. A person reads the intent, makes the calls and checks the result against the real thing. The failure log is what makes the numbers worth trusting, so it stays public.
Designing a precision measurement kit as a single system: one carrier, one housing language, one readout. All images are concept renders.
Most toolboxes are a collection of strangers.
A caliper from one maker, a micrometer from another, a square from a third. Each one is fine on its own. Together they share nothing: different housings, different screens, different buttons, and digital readouts fitted onto shapes that were designed before screens existed.
Datum 7 asks what a measurement kit looks like when it is designed as one object. Seven instruments, one carrier, one housing language and one readout. The name comes from the count.
It started as blocks: plain slabs in black and white, sized to the instruments, with no screens, keys or fasteners. Once the sizes felt right, the blocks became the first shells, with a display window, a key row and four corner screws. Laying every instrument out on one surface came next, and that layout turned straight into the first carrier.
Everything stayed one flat grey at that stage. Those renders were about proportion and order, not finish.
The angle finder decided the housing for the whole kit, and it took nine CAD concepts. Concepts 07 and 08 were a soft shell: rounded edges, a recessed grip, no visible fasteners. Concept X turned it into an instrument: a flat bezel held by four exposed corner screws, a two-key row under the display, and a three-point contact pad on the back.
That screwed shell, corner fasteners and key row then carried over to every other instrument. The hand tools have the same flat faces and chamfers as the digital ones, so the kit reads as one set from any distance.
The three digital instruments share one dot-matrix numeral set. Status words only light when they are active, so the display shows the unit and the mode without a menu. The caliper and micrometer get four keys: PWR, UNIT, DATA, ZERO. The angle finder only needs two: PWR and LOCK.
I rendered the caliper with a dead screen first, so the shell had to hold up with no graphics at all. The readout went in three days later.
Before the finish was set, every housing was rendered in four shell treatments: white, black, aluminium and smoked translucent. The test was legibility as much as looks, because the display has to read against every shell. The kit landed on machined aluminium, matte.
Every instrument has a fitted pocket in the one carrier.
It is not built yet, and the page says so. The accuracy figures on the back labels are design targets, not measured results, and the certification marks are layout placeholders. The carrier material and how it gets made are still open.
Every job on this site starts with a measurement. Datum 7 is the tool side of that loop: if measuring is the first step in making real things, the instruments deserve the same care as the parts they measure. Next is a working prototype, with the sensing, display driver and charging built for real.