Projects & Creations

Description

• Wrote a ROS2 driver for the Fanuc CRX-10iA that talks to the controller over EtherNet/IP, in C++ on top of the EIP Scanner library, so the whole cobot can be driven from ROS2 rather than the teach pendant.
• Built the operator UI in PyQt5 for the Fanuc CRX-10iA / Lincoln R450 PowerWave cell, and wrote a second driver in Python that speaks ArcLink over TCP/IP to the R450 power source for programmatic parameter control.
• The controller runs a set of background TP and KAREL programs I wrote that act as a thin firmware layer between the hardware and the ROS2 nodes.
• Now putting an AI agent into the live welding loop — LLMs and VLMs first for the high-level decisions, with smaller task-specific models to follow.

Technologies

Description

• Directed studies project (Fall 2026) on the same BCIT welding cell: measuring the geometry of a weld bead as it is laid down and using that to drive the process, all on ROS2 with the Fanuc CRX-10iA and Lincoln R450.
• A camera mounted on the robot hand (eye-in-hand) watches the bead. An existing width-estimation module is validated against a Revopoint Metro X Pro laser scan for ground truth, and I am adding a new height estimator that triangulates the bead surface from multiple frames using the known camera motion.
• The perception data feeds a trained feed-forward model — a learned inverse of the process — that takes a target bead width and height and predicts the travel speed and wire-feed speed needed to produce it, with a WPS bounds check clamping anything out of range before it reaches the controller.
• Part of the work is a feasibility study on whether the perception loop can run on the edge: profiling acquisition, processing, and end-to-end response time on the workstation first, then on CM4-class boards (Raspberry Pi, Radxa).

Technologies

Description

• Hardware design work for the BCIT welding-research vision system.
• The main piece is a custom flexible printed circuit (FPC): a two-layer flex cable that carries the camera interface between a Gowin GW1NR-LV9 FPGA board and the imaging sensor. It fans a 35-pin flip-lock connector out to a 23-pin one, on fine-pitch gold-plated contacts, and I had it fabricated through JLCPCB after a couple of order-review rounds on the flex stackup.
• Alongside it I designed a two-layer STM32F103 (Cortex-M3) MCU board — USB Micro-B, an AMS1117 3.3 V regulator, a 16 MHz crystal laid out to ST's AN2867 oscillator guidelines, boot-mode switches, status LED and an SWD header for ST-Link programming.

Technologies

Description

• Built with the team at the SFU Motion and Power Electronics Control Lab. The mechanical design came from the project lead; I took on the control software, the embedded systems, and the seam-tracking, and co-led the wiring and assembly.
• Prototyped the control software on a TMS320F240 DSP under dSPACE, then moved the final build onto a Raspberry Pi Compute Module 4 and an NXP S32K144 MCU.
• For seam tracking I paired the line-laser sensor with Python that filters the point cloud with KNN and pulls the edge out with piecewise-linear fitting, tight enough to correct in real time.
• On the theory side I worked out the DH kinematics, the analytical Jacobian, and the Lagrangian dynamics (inertia tensors from SolidWorks), then built PI and computed-torque controllers for the DC and PMSM motors and a jerk-minimised piecewise-polynomial trajectory generator that interpolates on the MCU. All of it was checked on dSPACE with Hardware-in-the-Loop before the portable build went together.
• The finished arm holds ±0.13 cm in height and 0.4° in rotation, with seam tracking and multi-pass welding emulation both proven on the physical robot. I also built a FreeMaster HMI and a browser dashboard over JSON-RPC (HTML/CSS/JS) for live control and MCU tuning.

Technologies

Description

• I was the FEA lead on a Level 3 Fitness-For-Service assessment (API 579-1 / ASME FFS-1, Part 8) of the 1912 riveted and welded firebox on the Spirit of Kamloops, Locomotive 2141. The report went to a regulator and was independently reviewed.
• Ran seven elasto-plastic simulations in SimScale over four versions of the geometry — the design intent, the as-measured distorted shape, and two weld-overlay repair options — under pressure-only and coupled transient thermo-mechanical loads at 200 psi MAWP.
• The as-measured firebox came from a Revoscan laser-scan point cloud, which I sliced, thickened to the real plate thickness, and swept into a clean parametric CAD model. Material models were bilinear isotropic hardening with temperature-dependent properties at the 350 °C design metal temperature, plus a separate E7018 weld-overlay material with its own calibration.
• To get the structural behaviour right I used symmetry planes, Winkler elastic supports with a per-stay stiffness K = EA/L tied to each missing stay's length, frictionless sliding contacts for the ball-end stays, and a six-zone heat-flux condition that blends convection and radiation from the 1500 °C fire side to the nucleate-boiling water side.
• Wrote a Python post-processor (PyVista/SciPy) to handle the stress and strain checks — projecting the assessment paths onto the deformed mesh, walking Dijkstra geodesics for the worst orientation, and running the local-strain (εL) and stress-linearization checks from ASME VIII Div 2 Annex 5.A. It agreed with the solver's own output to within ±2 ksi, and a 7× mesh-refinement study came back within ±5% when reproduced in Abaqus.
• Also wrote the internal report on the firebox corner-bend geometry, with an interactive MATLAB tool for plane slicing, PCA profile ordering, bend-angle and inscribed-arc extraction, and stay-pitch measurement.
• The result: the boiler was cleared to return to service at its full historical 200 psi MAWP with no derate, which avoided a repair estimated north of $200,000.
• Alongside the FFS work I ran cobot welding trials at BCIT, laser-scanning parts before and after to catch distortion, and built frontend tools for the welding engineers in JavaScript plus the website and database for a Fleet Management module in Ruby on Rails.

Technologies

Description

• Built a conveyor sorting system end to end in Studio 5000 on a CompactLogix L36ERMS Safety PLC.
• Kept the ladder logic modular — separate subroutines called through JSRs, with a state machine driving the sequence rather than one long rung stack.
• Wired in photo-eye sensors, pushbuttons, indicator lamps, and pneumatic actuators, and handled the timing with TON timers and CTU counters so the motion stays deterministic.
• Ran the PowerFlex 525 VFD off digital I/O for safe start/stop, and wrote fault logic on top of the timers and counters to catch missing bins, initialization timeouts, and E-stop conditions.

Technologies

Description

• Wrote the complete ladder logic control program for a continuous bottling line on Allen-Bradley SLC 500 addressing, developed and validated on the LogixPro simulator, and verified rung by rung against all 25 functional requirements.
• The governing design decision was to measure each bottle once at the inspection station and carry its properties downstream in software, so every later station acts on remembered data rather than re-sensing. Two parallel 32-bit BSL shift registers — one for size, one for condition — are clocked by the conveyor, so a bottle's flags stay aligned with its physical position at any bottle spacing.
• Fixed bit taps read each property at the exact position of the fill tube, scrap gate, cap RAM and divert gate. Property-gated actuation structurally enforces the safety rule that broken bottles are never filled or capped: the fill tube and cap RAM are gated by the tracked condition bit rather than a re-read sensor.
• A single on-delay timer (0.1 s base, preset 10) defines a 1.0 s processing window per bottle, with a LIM-gated 0.2–0.8 s actuation sub-window guaranteeing that fill, cap and gate motion always completes before the belt indexes.
• Five up-counters with TOD conversion multiplex four live production totals — small, large, scrap and completed boxes — through one shared BCD display word via dedicated enable bits, with a PREV BOX GONE / NEXT BOX HERE handshake sequencing the scrap-box changeover.
• Verification surfaced three residual fail-safe gaps that were not in the original requirements, all closed: fill charges re-gated to the master RUN bit so no charge persists through a Stop, a scrap-gate interlock against scrap-conveyor motion, and a first-scan (S:1/15) rung resetting every counter and timer to a known state on download.

Technologies

Description

• A mobile robot that navigates on its own with Hector SLAM and spots stop signs with a YOLOv3 model.
• On top of that, an OpenCV line-tracer keeps it on course and a small CNN reads cat and dog images off ArUco markers — the robot uses those cues to decide where to drop its red and blue balls and finish the mission.

Technologies

Description

• A research project comparing fiberglass-reinforced and steel-reinforced concrete under heavy loading, run in ANSYS.
• I looked at how deflection and stress distribution differed between the two, to see whether fiberglass rebar is a viable option for reinforcing concrete floors.

Technologies

Description

• Wrote my own ATmega328p library straight from the datasheet — registers, timers, ADC, PWM, interrupts — to get full use of the hardware without the Arduino layer in the way.
• Then built a PID speed controller on top and tuned it against the actual car.

Technologies

Description

• A technical review of K.C. Kong's work on Rotary Series Elastic Actuator control for human-robot interaction, with Simulink models I built to reproduce the control behaviour.
• The focus was compliance and force control, and what they mean for keeping physical human-robot interaction safe.

Technologies

Description

• A device that turns a flat 2D picture into a 3D touch sensation, aimed at helping blind and visually impaired people navigate. We proposed it and built a prototype.
• Deep learning pulls dense 3D depth out of the 2D image, and that depth gets rendered as touch on the user's skin through small electromechanical actuators.
• It ties a phone, an embedded system, and the deep-learning model together into a single wearable navigation aid.

Technologies