Muhammad Hussain Ahmad

Winnipeg, Canada

• Shipped a 1 kHz real-time C++ torque controller for a 7-DoF Kinova Gen3 with cyclic UDP and lock-free state, and implemented effort command handling and per-actuator torque mode switching in Kinova's ros2_control plugin.
• Resolved a 26.7 N·m torque error that no simulation reproduced by finding the Kinova API's torque sign reversed against the URDF and Pinocchio model, and traced a 198 ms torque-mode feedback gap to seven blocking mode-set round trips, proving it was not the fall cause with 0.07 N·m torque tracking.
• Built and tuned a 200 Hz Cartesian impedance controller on hardware with gravity and friction feedforward, null-space regulation, barrier terms in joint and task space, and 12 Hz torque filtering.
• Designed the actuator-level safety layer, with torque saturation and slew limits inside the 1 kHz loop and a 200 Hz supervisor for E-stop, watchdog, and position-mode handover, and replaced a latched-fault path that free-fell the arm 7.45 degrees with controlled fault handling.
• Added whole-arm self-collision avoidance over 26 URDF-mesh link pairs with barriers projected through Jacobians at 0.39 ms p99.
• Identified payloads online with an Extended Kalman Filter, 0.96 kg at a 97 mm wrist offset, holding joint residual within 0.57 mrad at 0.45 ms median per step over 448k live steps, while an offline baseline never recovered.
• Predicted contact 198 ms ahead at 100% TPR with a Genesis digital twin, holding the 10 ms real-time budget at the 100 ms horizon (6.3 ms median, 8.0 ms p99).
• Held a 15-node 1 kHz to 100 Hz ROS 2 stack on schedule after the OS denied SCHED_FIFO by sequencing bring-up around a 250 ms DDS discovery stall and capping the watchdog below 0.25 s.
• Cut end-effector tracking error 57% on average (43% vs RelaxedIK, 59% vs TRAC-IK, 71% vs Cartesian impedance) over 1,080 markerless camera teleoperation trials with a CasADi QP-based numerical IK retargeter at 2.59 ms p99 and zero missed deadlines.
• Engineered a camera-only intent channel with dwell-time gating, cutting false grip activations to 0.16 per minute across 195 minutes by discarding 97.3% of 1,323 spurious requests.
• Developed a browser-based (HTML and JavaScript) hardware debugging console with live torque, tracking, fault, and mode-control telemetry.

Winnipeg, Canada

• Operated a ROS 2 scanning pipeline on a Unitree Go2 EDU's onboard Jetson Orin for daily 4-hour runs on a construction site, with a Leica BLK360 and Insta360 on a custom 3D-printed mount.
• Delivered Scan-to-BIM point clouds at 95% export success by clearing robot, sensor, and embedded-compute failures on site.

USA (Remote)

• Owned PyTorch inference for Stable Diffusion XL, reducing end-to-end latency from 2.5 s to 0.7 s with LCM-LoRA, execution-order changes, and GPU scheduling.
• Tripled 1024×1024 generation throughput via dynamic batching and VRAM tuning, cutting per-step compute time by 30%.
• Cut Flux VRAM from 28 GB to 13 GB with BF16 and TensorRT.

Remote

• Shipped a virtual staging pipeline with SDXL, ControlNet, and Depth Anything, and cut its Docker builds from 2.5 min to 40 s.
• Owned production Llama 3 8B inference for Optimization Ventures, cutting latency by 35% with INT8 quantization and execution-graph optimization while preserving model accuracy.
• Improved serving stability under concurrency by reducing p95 latency variance by 40% through execution-path cleanup and removal of redundant ops in the inference pipeline.
• Trained a Siamese face-verification CNN in TensorFlow with L1 distance scoring and a real-time OpenCV webcam verification pipeline.
• Trained a hybrid CNN + Vision Transformer on 1M images for 7-class emotion recognition, achieving 90% top-1 accuracy and improving mAP by 12% over a CNN baseline, with 18% higher minority-class F1 from focal-loss tuning and targeted augmentation.

Owned evaluation harness that catches silent checkpoint regressions before deployment by scoring OpenVLA / SmolVLA / diffusion / scripted policies behind one adapter and reporting failure rate across 7 perturbation axes instead of aggregate means.

• Diffs each checkpoint against the last on a fixed benchmark so regressions surface as axis-wise failure reports, not a single success-rate number.
• Parallel rollout runner and plugin system over 4 simulators (MuJoCo, PyBullet, Genesis, Isaac), OpenVLA & SmolVLA adapters, ROS 2 publishing, and 3,678 pytest cases gated per PR on GitHub Actions.
• Shipped to PyPI with a full write-up and architecture diagrams.

1 kHz real-time C++ torque controller for a 7-DoF Kinova Gen3 over cyclic UDP with lock-free state, integrated through effort command handling and per-actuator torque mode switching in Kinova's ros2_control plugin. A 200 Hz Cartesian impedance controller and an actuator-level safety layer (torque saturation, slew limits, E-stop, watchdog, and position-mode handover) run on top, tuned and validated on the physical arm.

End-to-end markerless RGB-D teleoperation in ROS 2 driving a 7-DoF Kinova Gen3 from MediaPipe pose estimation. Perception, retargeting, safety, and command lanes run at 100 Hz with a 2.59 ms p99 control tick over 30k+ live ticks and zero deadline violations.

CasADi QP retargeter with null-space posture shaping that cut end-effector tracking error by 57% on average: 43% vs RelaxedIK, 59% vs TRAC-IK, and 71% vs Cartesian impedance across 1,080 trials (also vs DexPilot-style).

Predictive safety gate driven by a Genesis digital twin that caught 100% of simulated contact events a mean 198 ms early, while holding the 10 ms control budget at the 100 ms horizon (6.3 ms median, 8.0 ms p99). Online EKF and moving-horizon correction keep residual within 0.57 mrad after payload changes, identifying a 0.96 kg payload at a 97 mm wrist offset.

Designed and built a 75 kg linear load cell, bonded its strain gauges, and calibrated it with Arduino, using SolidWorks and ANSYS FEA (134.76 MPa peak stress, 965,000-cycle fatigue life). An S-beam variant was designed and analyzed in simulation.

Production PyTorch inference optimization for SDXL at Wombo: cut end-to-end latency from 2.5s to 0.7s and raised 1024×1024 throughput by 3× via LCM-LoRA, dynamic batching, VRAM tuning, and attention-path improvements.

Hybrid CNN + Vision Transformer trained on 1M images for 7-class emotion recognition — 90% top-1 accuracy, +12% mAP over a CNN baseline, and +18% minority-class F1 via focal loss and targeted augmentation.

Self-gated hierarchical Transformer for 21-class fault classification on the Tennessee Eastman Process benchmark, with gate-based sensor attribution and a Transformer autoencoder for unsupervised anomaly detection. A follow-on with a collaborator adds class-imbalance-aware training (margin head, contrastive and class-centre losses, mixup).