FULL VERTICAL SLICE. I BUILD THE ↑

Give it a CAD part.
It welds it.

Built for a cobot welding integrator. Customer part data and commercial terms withheld — described at the level of problem and approach.

It finds the weld seam, works out the toolpath, checks whether the arm can reach it without burying its own elbow in the workpiece, aligns the path to where the part really sits, and drives a Dobot CR10A along it.

• The brief was one sentence. Put a camera on the arm, let it pick the points to be welded, then weld them. Everything after that sentence — research, stack selection, architecture, a thirteen-phase roadmap, implementation — was mine. There was no reference implementation. 36,321 lines of source, 28,409 of tests, 145 commits in about three weeks.
• Geometry, not guesswork. STEP parsing through OpenCASCADE, seam detection by concave-edge analysis, joint-type classification (fillet / butt / lap) from opening angle and face arrangement, then oriented 3D waypoints with lead-in and lead-out. Rule-based by choice — ML only where rules genuinely cannot reach.
• I solved the arm myself. A closed-form inverse kinematics solver for the CR10A, and a body-clearance model whose link capsules are measured off the robot's own mesh instead of estimated. That measurement caught a wrong driver constant and a prior model drawing the upper arm 613 mm oversized. Reach is checked against the whole body of the arm, not the tool tip.
• The part is never where the CAD says. A depth camera and ICP registration recover the real pose and correct the CAD-derived path. A placement sweep then answers "where can this part stand" as a band of valid positions rather than one yes/no — and the band narrows the moment a 200 mm torch is fitted.
• 29 July 2026 — first run on a real CR10A. Connected, enabled, drove a full path in real motion. Motion only: no torch fitted, no arc struck. It stopped once mid-run at a singularity, so the next version got a pre-flight singularity check. A fake controller was written so the driver could be developed without hardware, then given the real closed-form CR10A IK.

TARGET HARDWARE

Dobot CR10A

SOLVER

Closed-form inverse kinematics, written for this arm

CLEARANCE

Link capsules measured off the robot's own mesh

FIRST RUN

29 July 2026 — connected, enabled, drove a full path

ARC

No torch fitted, no arc struck. Motion only.

Joints you can swap while it is running.

Instead of a fixed kinematic chain baked into a URDF at build time, the arm discovers its own configuration and rebuilds its kinematic description at runtime. Pull a limb, put a different one on, and it re-derives the chain without a restart.

• Joint modules are hot-swappable while the system is live.
• ROS2 + MoveIt2 pipeline operational and demonstrable in RViz.
• Distributed joint control on bare-metal STM32 over CAN — each module is an addressable node running its own control loop.
• Where it is going: AI-driven self-reconfiguration — the arm choosing its own morphology for a given task.

A platform that refuses to fall over.

A real-time self-balancing platform built from the driver up: closed-loop PID over fused IMU data, driving brushless motors through six-step commutation — including the unglamorous part, bringing up the motor driver, finding the noise, and tuning it on real hardware rather than in simulation.

• Six-step commutation with rotor position feedback; phase timing tuned against measured back-EMF behaviour.
• Sensor fusion to get a usable attitude estimate out of a noisy IMU before the controller ever sees it.
• Control loop written bare-metal — no RTOS, deterministic timing, measured on a scope rather than assumed.

The numbers are computed, not generated.

Built in a two-person team with AI-assisted development under enforced quality gates. Client data, pricing logic and internal rules withheld; the architecture below is a general pattern.

A system that reads incoming automation project documents and produces a first-pass technical feasibility assessment — the work a senior engineer does in the first two hours of looking at an enquiry.

• Payload, reach, torque, cycle-time and cost checks run as deterministic Python functions the model calls as tools. The LLM interprets the results; it never invents them.
• Hybrid retrieval — vector plus lexical, reciprocal rank fusion — over a structured 46-table schema, scoped by process and sector rather than keyword-matched.
• Every claim carries a reference back to a record. Unsourced statements are labelled engineering estimates, and commercial data is stripped from the response server-side rather than hidden in the UI.
• Human-in-the-loop by design: nothing is released without engineer approval, and approved corrections return as reusable knowledge.