My Journey
Professional Experiences
A timeline of my professional growth, from teaching and leadership roles to hands-on technical experience in robotics and software development.
Tech
Research Associate
BCIT Centre for Welding Technologies and Metallurgy Research
2025 - PresentBurnaby, BC
Highlights
- Derived the rigid-body dynamics of the Fanuc CRX-10iA from its URDF (recursive Newton–Euler and composite rigid-body algorithms, written from scratch) and designed six joint-space controllers on it: computed torque, MPC as a constrained QP, ILC, ADRC, and learned and adaptive inverse-dynamics control. Compared them over 3 studies × 12 randomised weld trajectories: computed torque and ILC reach 0.013° RMS on the nominal model, the adaptive controller holds 0.077° under an unmodelled payload where computed torque and MPC degrade to 2.6–3.1°, and MPC does best under torque limits.
- Modelled the cross-track seam-tracking loop with its transport delay and derived the stability boundary from z^(d+1) − z^d + K = 0 (K < 0.618 at two segments of delay). The shipped gain of 0.6 was on the boundary, which explained the ringing. Designed and compared delay-compensating estimators (Smith predictor, trajectory buffer, Kalman), which cut worst-case error from 1.30 mm to about 0.8 mm in the plant model, then tested them in a 20-weld interleaved campaign in Isaac Sim.
- Built a physics-based model of the weld process: a Goldak double-ellipsoid heat source and Rosenthal moving-source temperature fields (t8/5 = 8.1 s at the nominal pass, inside EN 1011-2's band), the arc solved as a constant-voltage circuit from Lesnewich's melting rate, inherent-strain distortion with a beam model, weld-pool Marangoni flow, and a Young–Laplace bead cross-section.
- Showed with that model that the disturbance a seam tracker actually faces is a 5.4 mm ramp over a 300 mm fillet pass, not the 3 mm step it had been tested against, and that the deposited bead accounts for 83% of the fillet's angular distortion.
- Turned the models into a procedure tool with an eight-clause envelope (heat input, t8/5, burn-through, transfer mode, AWS D1.1 leg size, CTWD, torch angles). It caught the cell's own shipped parameters running in the globular band at 9.6% spatter, with t8/5 below the standard's floor on 12 mm plate.
- Brought the cell up in Gazebo and NVIDIA Isaac Sim 6.0 from one shared SDF/URDF, with overview-camera joint detection and torch-camera seam tracking, so controllers and perception run unchanged in either simulator.
- Modelled spatter, fume and arc sound from the arc's own operating point (sound as a thermoacoustic monopole, p ∝ dP/dt, driven at the metal-transfer rate) and a ray-traced laser line profiler that measures the simulated bead the way a real one would, occlusion included.
- Building a vision pipeline (directed studies, Fall 2026) that reads weld-bead width and height off an eye-in-hand camera in real time, validated against Revopoint laser scans, plus a trained feed-forward model that picks the travel and wire-feed speeds needed to hit a target bead shape.
- Built the cell's software interfaces: a C++ ROS2 EtherNet/IP driver for the Fanuc (EIP Scanner), a controller-side job dispatcher and job set in TP and KAREL, a C++ ROS2 driver for the R450 over Lincoln's ArcLink/XT protocol (ported from my original Python), and the PyQt5 operator panel with live arc data, jogging and the top-camera view. Every pass is logged with its current, voltage and wire-feed traces to the lab's Rails/PostgreSQL weld database.
- Added a local voice agent to the real cell (faster-whisper plus a 4-bit Qwen3-4B tool-calling model, no cloud service). Its tool calls are checked against what was actually heard and against the cell state, and risky commands such as striking an arc need a spoken confirmation.
- Mapping unknown workpieces with the welding wire's touch sense (paper in preparation, "Look, Then Touch"): a step-and-bracket contour tracker plus a measured error budget on the cell. Contact height repeated to 0.02–0.06 mm s.d., the wire drifted about 7.5 µm per touch, doubling the search speed shifted it by about 25 µm, and four touched edges placed a rotated plate within 0.18° of ground truth. Touch-anchored correction of a RealSense D435i's depth with a Kalman filter is in progress.
- Now working on putting an AI agent inside the welding loop — LLMs and VLMs first for the high-level reasoning, and eventually smaller task-specific models trained just for the welding steps. Alongside it, built a physical-safety benchmark for AI-written welding programs that scores the executed trace against a welding hazard taxonomy.
Skills
System ModelingRobot DynamicsControl DesignMPCStability AnalysisTouch SensingSensor FusionKalman FilterWeld Process ModelingHeat TransferROS2C++PythonIsaac SimGazeboSeam TrackingEtherNet/IPArcLink/XTComputer VisionLLM/VLMAI Safety
Junior Systems Engineer
SKC Engineering Ltd
2025 - PresentVancouver, BC
Highlights
- Ran the FEA side of a Level 3 Fitness-For-Service assessment (API 579-1 / ASME FFS-1, Part 8) on the 1912 riveted and welded firebox of a locomotive boiler. The report was regulator-facing and independently reviewed.
- Set up and ran seven elasto-plastic simulations in SimScale across four versions of the geometry — the original design intent, the as-measured distorted shape, and two weld-overlay repair options — under both pressure-only and coupled transient thermo-mechanical loads at 200 psi MAWP.
- Rebuilt the as-measured firebox from a Revoscan laser-scan point cloud: sliced it, thickened it out to the measured plate thickness, and swept it into a clean parametric CAD model that would actually mesh.
- Built the material models myself — bilinear isotropic hardening with temperature-dependent conductivity and expansion, evaluated at the 350 °C design metal temperature — plus a separate E7018 weld-overlay material with its own calibrated properties.
- For the boundary conditions I used symmetry planes, Winkler elastic supports with a per-stay stiffness of K = EA/L tied to each missing stay's length, and frictionless sliding contacts for the ball-end stays with heat transfer across the contact.
- Applied a six-zone convective heat-flux thermal condition blending convection and radiation, from the 1500 °C fire side to the nucleate-boiling water side, consistent with saturation conditions at the design pressure.
- Wrote a Python post-processor (PyVista/SciPy) that projects the assessment paths onto the deformed mesh, walks Dijkstra geodesics to find the worst orientation, and runs the pointwise local-strain (εL) and through-thickness stress-linearization checks from ASME VIII Div 2 Annex 5.A. It matched the solver's own post-processing to within ±2 ksi.
- Ran a mesh-refinement study at 7× the node count to confirm the staybolt-anchor stress concentrations were captured conservatively, then had the results reproduced independently in Abaqus to within ±5%.
- Wrote the internal report on the firebox corner-bend geometry from the scan, with an interactive MATLAB tool for plane slicing, PCA profile ordering, bend-angle and inscribed-arc extraction, and stay-pitch measurement.
- The assessment cleared the boiler to go back into service at its full historical 200 psi MAWP with no derate, which avoided a repair estimated north of $200,000.
- On the welding side, I ran cobot welding trials at BCIT and laser-scanned the parts before and after to measure distortion from the point clouds.
- Also built frontend tools for the welding engineers in JavaScript, stood up the website and database for the crane-certification module of a Fleet Management system in Ruby on Rails, and handled CAD drafting for deliverables.
Skills
FEAAPI 579-1 / ASME FFS-1ASME VIII Div 2SimScaleANSYSAbaqusElasto-Plastic AnalysisThermo-Mechanical FEAPython (PyVista/SciPy)MATLABSolidWorksAutoCADLaser ScanningPoint CloudJavaScriptRuby on Rails
Graduate Research Assistant
SFU Motion and Power Electronics Control Lab
May 2022 - Dec 2024Surrey, BC
Highlights
- Helped build an in-pipe welding robotic arm with the lab team. The mechanical design came from the project lead; I took on the modeling and control, the embedded systems, and the seam-tracking algorithms, and co-led the electrical wiring and assembly.
- Worked out the full kinematic and dynamic model: Denavit–Hartenberg forward kinematics, the analytical Jacobian, workspace analysis, and closed-form inverse kinematics that I solved two separate ways and cross-checked in Simulink.
- Derived the rigid-body dynamics with the Lagrangian method, pulling the inertia tensors from SolidWorks, and used that model to design a computed-torque controller, tuned with frequency-response and time-domain methods, which I benchmarked against a simpler independent-joint PI speed controller.
- Designed a trajectory generator for smooth, jerk-free motion, interpolated in real time on the MCU.
- Prototyped everything on dSPACE with Hardware-in-the-Loop testing in Simulink before committing to the portable build, where the Pi CM4 handles the laser processing and the S32K144 runs the real-time motor control, the two linked over SPI.
- Measured closed-loop accuracy came out to ±0.13 cm in height and 0.4° in rotation, with seam tracking and multi-pass welding emulation both validated on the physical robot.
- Started 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 ARM MCU.
- Built the motor drive stage: BTS 7960 dual H-bridge drivers for the six DC arm axes (43 A, PWM speed and direction, with current-alarm handling) and a SOLO UNO v2 controller for the BLDC rotation axis.
- Got a Wenglor weCat3D MLSL143 3D profile laser sensor talking over raw TCP/IP sockets — the ASCII command handshake, encoder-mode scanning, and unpacking the FPGA-packed binary profile frames in Python — and fed that into a live seam-tracking loop.
- Used KNN for point-cloud filtering and sharpened the edge detection with piecewise-linear fitting tuned by differential evolution.
- Built a FreeMaster HMI and a browser dashboard (HTML, CSS, JavaScript, JSON-RPC) for real-time control and MCU parameter tuning.
Skills
DH KinematicsLagrangian DynamicsComputed-Torque ControlPI ControlTrajectory GenerationSimulinkdSPACE / HILMATLABC++PythonROS2Raspberry Pi CM4NXP S32K144TMS320F240 DSPSPITCP/IPKNNDifferential Evolution3D Profile Laser SensingJSON-RPCOpenCV
Leadership
Teaching Assistant — Systems Modeling, Control & Mechatronics
Simon Fraser University
Sep 2022 - Dec 2024Surrey, BC
Highlights
- TA for MSE 380 Systems Modeling (Fall 2022, 2023), MSE 483 Modern Control Systems (Spring 2023, 2024), MSE 310 Sensors and Actuators (Fall 2024), MSE 312 Mechatronics Systems Design 2 (Summer 2023) and MSE 112 Mechatronics Design Studio (Summer 2024).
- Wrote new lab and lecture material for the Mechatronics Design Studio (MSE 112), based around a robot kit with Mecanum wheels, a 4-DOF arm, a Raspberry Pi 5, and an MCU expansion board.
- Ran the lab sessions: students working with motors and motor drivers, waveform generators, oscilloscopes, and power supplies.
- Helped them get comfortable with Simulink, LabVIEW, Python, and C++ for real-time simulation and embedded programming.
- Spent a lot of the time one-on-one — debugging their code, chasing down hardware issues, and keeping projects moving.
Skills
System ModelingModern ControlTeachingSimulinkLabVIEWPythonC++MechatronicsMentoringCurriculum Design
Education
PhD student, Mechatronic Systems Engineering
Simon Fraser University
Jan 2026 - PresentBurnaby, BC
Highlights
- Research sits at the intersection of Physical AI and robotic welding, run jointly with BCIT's Centre for Welding Technologies and Metallurgy Research.
- Currently working through the qualifying exam. Awarded the PhD Research Scholarship for Spring and Summer 2026.
Skills
Physical AIRobotic WeldingLLM/VLM AgentsResearch
Master of Applied Science (MASc), Mechatronic Systems Engineering
Simon Fraser University
May 2022 - Jun 2025Burnaby, BC
Highlights
- Thesis: Development of an In-Pipe Welding Robotic Arm for Seam Tracking and Multi-pass Welding (completed December 2024, degree conferred June 2025).
- GPA 4.17 / 4.30.
- Awarded SFU Graduate Fellowship (2023 and 2024) and the SFU Graduate Entrance Scholarship (2022).
- Relevant coursework: CMPT 726 Machine Learning; CMPT 983 Trustworthy Deep Learning.
Skills
RoboticsControl SystemsEmbedded SystemsMachine Learning
Graduate coursework — Robotics, Automation and Industrial Control
Concordia University
Aug 2021 - May 2022Montreal, QC
Highlights
- Graduate coursework in industrial control systems and robotics; 6 credits (MECH 6621 Microprocessors & Applications, MECH 6021 Design of Industrial Control Systems) transferred into the SFU MASc. Not a completed degree.
- Also completed ENGR 6411 Robot Mechanics and MECH 6631 Industrial Automation.
Skills
Industrial ControlMicroprocessorsRobot Mechanics
BSc, Mechanical Engineering
Korea Advanced Institute of Science and Technology (KAIST)
Feb 2017 - Aug 2021Daejeon, South Korea
Highlights
- Full KAIST Scholarship recipient (2017).
- Relevant coursework: ME440 Engineering Design via FEM; ME460 Automatic Control; ME491/492 Autonomous Mobile Systems; CoE202 Big Data Analysis & Machine Learning.
Skills
Mechanical DesignFEMAutomatic ControlAutonomous Systems