All work

Computer vision + manufacturing automation · 2026

PatchSight MK2

PatchSight is the production side of the PATCH22 ecosystem. It takes a physical patch, finds a usable contour, converts pixels into millimeters, and asks an operator to review the result before anything leaves the system.

CALIBRATION VALID · CONTOUR CLOSED · OPERATOR REVIEW REQUIREDFRAME / 014
Capture, validate, review, then export. The operator remains in control.Derived from verified PatchSight MK2 vision and calibration boundaries; no proprietary CAD used.
My role
Product and systems engineering
Status
Stationary-bed MVP
Contribution
Vision pipeline, calibration rules, operator review, and safe export boundaries.
01

Here’s the problem

The work has to make sense before the interface can.

An edge that looks plausible on camera is not automatically a safe cut path. Calibration, geometry, and operator authority all matter.

02

What made this difficult

The environment is part of the product.

  • Camera and calibration failures must remain visible.
  • Open, invalid, or out-of-bounds geometry must fail closed.
  • The system cannot control the laser; an operator approves and exports the file.
03

Decisions that shaped it

I made the hard boundaries visible.

Separate the boundaries

Capture, vision, geometry, calibration, jobs, and export are explicit parts of the system.

Make review unavoidable

The detected contour is shown over the source image before an operator accepts or rejects it.

Keep machine authority out

The MVP exports calibrated SVG geometry for manual import instead of driving the laser directly.

04

What I built

Enough of the system to test the real seam.

  • Python 3.12 engine using DepthAI, OpenCV, Shapely, NumPy, and Pydantic.
  • Native SwiftUI operator surface calling the local engine through a constrained JSON CLI.
  • Immutable job artifacts and millimeter-based SVG export.

What is true now

A reviewable path from camera capture to calibrated SVG output, without giving the software authority over the laser.

Limits I’m keeping honest

  • The MVP assumes a stationary work bed.
  • Cut quality and machine execution remain outside the application’s verified authority.

Next

Validate repeatability across more patch shapes, camera conditions, and calibration intervals.

Next case study

PipeTrace

A problem like this hiding in your operation?

Bring me the difficult part.

Start a conversation