Muhammad Hussain Ahmad

M.Sc. Mechanical Engineering (Robotics) @ University of Manitoba

Winnipeg, MBmhussainahmad@outlook.com+1 (204) 298-7998

MHA

Hi! I am a Machine Learning Engineer with 4+ years of experience building production ML systems for perception, real-time control, and high-performance inference. Strong Python and PyTorch background across RGB-D perception, sensor fusion, transformer and diffusion-based models, distributed training, and GPU optimization. Experienced in simulation-driven validation, large-scale model evaluation, and deploying perception-to-control pipelines on physical robotic systems.

Experience

  1. University of Manitoba · Research Assistant (Robotics & Perception)

    May 2024 – Present

    Winnipeg, Canada

    • Owned the end-to-end markerless RGB-D teleoperation stack in ROS 2 that drives a 7-DoF Kinova Gen3 from MediaPipe pose estimation, running at 100 Hz with a 2.59 ms p99 control tick over 30,203 live ticks against a 10 ms limit, with zero violations.
    • Designed a CasADi QP retargeter with null-space posture prior that cut end-effector tracking error by 43% vs RelaxedIK, 59% vs TRAC-IK, and 71% vs Cartesian impedance (p < 10⁻²⁰), evaluated over 1,080 trials against TRAC-IK, RelaxedIK, DexPilot-style, and Cartesian impedance.
    • Engineered a camera-only intent channel with dwell-time gating, cutting false grip activations to 0.16 / min across 195 minutes by discarding 97.3% of 1,323 spurious requests.
    • Caught 100% of simulated contact events a mean 198 ms early with a Genesis digital twin at a 200 ms horizon, while holding the 10 ms real-time budget at the 100 ms horizon (6.3 ms median / 8.0 ms p99).
    • Kept the twin accurate under payload changes with an online sim-to-real EKF on the physical arm at 0.45 ms median / step (448k live steps), holding joint residual at ≤ 0.57 mrad after payload add while an offline baseline never recovered.
    • Kept torque-mode interaction fail-safe with a reactive last-resort chain that halts within 260 ms of a fault, behind the predictive twin gate.

  2. Wombo · Machine Learning Engineer

    Aug 2024 – Dec 2024

    USA (Remote)

    • Owned PyTorch inference for Stable Diffusion XL, reducing end-to-end latency from 2.5s → 0.7s through execution-order changes, GPU scheduling, and memory-allocation tuning.
    • Increased 1024×1024 generation throughput by 3× via dynamic batching and VRAM optimization, cutting per-step compute time by 30%.

  3. Self-employed · Freelance Machine Learning Engineer

    Jan 2022 – Jul 2024

    Remote

    • Owned production autoregressive inference for Optimization Ventures, cutting latency by 35% using 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 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.
    • Improved minority-class F1 scores by 18% using focal-loss tuning and targeted augmentation.

Research Interest

My research and engineering work focuses on RGB-D perception, real-time robot control, sim-to-real transfer, and high-performance ML inference. Recent focus areas include:

  • Gauntlet — catches silent checkpoint regressions across MuJoCo, Genesis, Isaac Sim, and PyBullet with axis-wise failure breakdowns over 7 perturbation axes
  • Markerless RGB-D teleoperation and pose-to-joint retargeting for manipulators (Kinova Gen3, MediaPipe, RealSense)
  • Constrained optimization and safety-gated control stacks in ROS 2
  • Digital twins and predictive safety with Genesis / NVIDIA Isaac Sim
  • GPU-optimized inference for diffusion and vision models in production

Publications

  1. A Component-Wise Ablation Study of Real-Time Audio-Visual Emotion Recognition

    Submitted to GLOBCER

    Submitted
  2. A Data-Driven Group-Level Recommender for Object and Difficulty Selection in Cerebral Palsy Rehabilitation Games

    Submitted to JMIR Rehabilitation and Assistive Technologies

    Submitted
  3. Subject-Coupled Markerless End-Effector Teleoperation for Therapist-Guided Arm Rehabilitation with a Constrained Optimization Retargeter
    In progress
  4. A Confidence-Weighted Multi-Modal Gesture Arbitration Framework for Markerless Rehabilitation Teleoperation
    In progress
  5. Predictive Simulation-in-the-Loop Shared Autonomy for Markerless Rehabilitation Teleoperation
    In progress
  6. Online Sim-to-Real Parameter Adaptation for Predictive Teleoperation: Bounded-Divergence Tracking of the Kinova Gen3 in Genesis
    In progress
  7. Subject-Specific Range-of-Motion Adaptation for Therapist-Guided Rehabilitation Teleoperation: Method and Evaluation Protocol
    In progress

Projects

  1. Gauntlet

    Details

    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.
    MuJoCoGenesisIsaac SimPyBulletOpenVLAROS 2Python

  2. RGB-D Teleoperation Stack

    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.

    ROS 2Kinova Gen3RealSenseMediaPipeC++Python

  3. Constrained Motion Retargeting

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

    OptimizationCasADiInverse KinematicsHRI

  4. Genesis Predictive Safety Twin

    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 / moving-horizon correction keeps residual ≤ 0.57 mrad after payload changes.

    GenesisSim-to-RealSafetyDigital Twin

  5. Stable Diffusion XL Inference

    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 dynamic batching, VRAM tuning, and attention-path improvements.

    PyTorchDiffusionGPU OptimizationServing

  6. Emotion Recognition (CNN + ViT)

    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.

    PyTorchTransformersComputer Vision

Education

  1. University of Manitoba

    M.Sc., Mechanical Engineering (Robotics Specialization) · Winnipeg, Canada

    May 2024 – Dec 2026 (Expected)

  2. National University of Sciences and Technology (NUST)

    B.S., Mechanical Engineering · Islamabad, Pakistan

    Sep 2019 – June 2023

Skills

PythonPyTorchROS 2C++NVIDIA Isaac SimIntel RealSense D455Kinova Gen3OpenCVMediaPipeRGB-D Perception3D Pose EstimationSensor FusionReal-Time ControlSim-to-RealTransformersDiffusion ModelsDistributed TrainingGPU OptimizationQuantizationDockerAWSCI/CD