From Phone Photos to a Part That Fits: Marine Reverse Engineering with 3DJ
Updated: Oct 8
A small marine replacement part became a practical test of how we build AI design tools: start with the problem, measure what matters, and test the result on the real object.
For Malta Yachting, Neural Deep Network LTD reverse-engineered an AutoAnchor windlass remote cradle from phone photographs. The work ran from 23–30 September 2026. After three design iterations, the v3 easy-snap version was printed in ASA and fit-tested on the actual handset on 30 September. It fits.
The brief: replace a part without a drawing
The cradle mounts a windlass remote to a bulkhead and holds it between two flexible side walls. Malta Yachting needed a replacement, but there was no drawing or spare part in stock. Our starting material was phone photographs, some with a ruler, plus product listings.
This is a familiar marine and refurbishment problem: the larger system still works, but one small component is unavailable. Recreating that component requires understanding its functional interface, not simply matching its silhouette.
Three iterations—from plausible to fitting
The first design was based on photographic proportions. At 64 mm across the arms, it was about 6 mm too wide. It looked plausible, but would not have fitted.
Ruler photographs brought the overall length to 88.5 mm, the width over the arms to 57.8 mm, and the arm height to 19.5 mm. The second iteration corrected the overall geometry, but its retaining lip was too aggressive.
Annotated close-ups then revealed details hidden from the top view: the step, undercut and lead-in around the lip. The third iteration incorporated that feedback. The printed easy-snap variant used a 1.3 mm lip undercut.
The useful lesson is simple: put the measuring effort into the interface that decides whether the part works.
A ruler gives a scale for one plane
Features raised above a ruler are closer to the camera and appear larger. In this case, the measured scale was 14.10 pixels/mm at the bench and 15.19 pixels/mm at the arm tops, 19.5 mm higher.
Using the bench scale for the raised lips would have over-read their size by 7.7%. Across a 47.8 mm opening, that is roughly 3.7 mm—larger than the lip itself.
Height-aware calibration helped us avoid that error. For symmetric details, measuring between mirrored features also reduced the effect of the part's rotation in the photograph.
Every dimension needs a source
We recorded whether a dimension was measured, traced from a photograph, estimated, or supplied by the user. Hidden details remained labelled as estimates until better views or measurements were available.
That makes the next step clearer for the customer: which measurements should be checked before committing to a full part or a batch?
Design the snap-fit, then test it
The retaining arms bend as the handset enters. Increasing the lip can increase retention, but also increases strain at the arm root.
A first-order snap-fit calculation helped compare the variants. The case study estimated about 3.4% root strain for v2, 2.5% for the standard v3 lip, and 1.9% for the printed easy-snap variant. The ASA allowance used in this calculation was derated for bending across FDM layers.
These are engineering screening estimates, not a durability certification. The real walls have additional constraints, and printed material behaviour needs physical confirmation.
We also designed a smaller fit coupon containing the critical arms and lips, with a window in the base. Its estimated volume was 7.25 cm³ compared with 12.92 cm³ for the full cradle—about 44% less material for a targeted interface test.
Independent checks before printing
The design story documents a point-by-point comparison between the v2 mesh generator and OpenSCAD description: zero mismatches in the sampled test. The STL files were also checked for a single watertight shell and manifold geometry.
Those checks answer different questions. Geometry verification helps catch broken or inconsistent models; the physical fit test tells us whether the model works with the actual handset.
The final ASA print was produced on a Bambu Lab A1 with a 0.4 mm nozzle, six walls and 85% infill. The reported filament cost was under €0.50 per part. That is material cost only, excluding design, machine time, labour and installation.
What this work added to 3DJ
The job fed reusable methods back into our 3DJ Studio:
Height-aware photo scaling and camera fitting, including a residual to expose inconsistent reference measurements.
Mirrored-pair averaging for rotation-independent spans.
Snap-fit strain checks to compare lip geometry and material assumptions.
Point-in-solid testing, independent geometry cross-checks and fit-coupon planning.
The supplied report records 38 passing checks and zero failures in the verification harness. These methods are in the codebase; connecting them to the Photo tab and the checks interface is described as the next development step, rather than a shipped user-interface feature.
The result—and the next evidence we need
On 30 September 2026, the v3 easy-snap cradle was tested on the real handset for Malta Yachting: it fits.
Long-term service on a boat, UV ageing and repeated insertion cycles have not yet been tested. A successful fit is an important milestone; those follow-up tests will tell us more about life in service.
For us, this is what AI design agents should support: faster iteration grounded in measurements, deterministic geometry and physical feedback. “AI plans, math builds.”
Have a broken or unavailable part?
Send clear photographs with a ruler in the frame, close-ups of the working interfaces, and a description of how the part is used. We can assess what is measurable, what remains uncertain and what a useful prototype or fit coupon should test.
Contact 3d@neuraldeep.net or explore 3DJ Studio.
This is an independent replacement design. AutoAnchor, Muir, Lewmar and Imtra names identify the original part; no OEM affiliation is implied. Case-study client: Malta Yachting.
Comments