• Developing physics-aware sensorimotor representation learning for embodied agents to improve sample efficiency in downstream control tasks.
• Designing and evaluating sim-to-real transfer pipelines (domain & dynamics randomization, latent-space alignment) for manipulation policies using PyBullet and real robot validation.
• Implementing modular data-collection tools and evaluation suites to measure robustness under visual and dynamics perturbations; coordinating hardware-in-the-loop experiments.
• Outcome: prototype sim-to-real policies, reproducible experiment scripts, and internal reports toward journal/conference submissions.

• Profiled DNN workloads across heterogeneous accelerators to derive time & energy rooflines and identify bottlenecks for edge deployment.
• Designed energy-aware model scheduling strategies and lightweight runtime policies to trade off latency, accuracy and power on resource-limited devices.
• Built measurement pipelines (benchmark harnesses, power telemetry, micro-benchmarks) and performed controlled experiments on representative CV and speech models.
• Outcome: empirical roofline analyses, scheduling prototype, and poster presented at an MLSys workshop.

• Researched deep reinforcement learning algorithms for dexterous manipulation tasks including grasping, in-hand rotation and contact-rich maneuvers.
• Developed curricula and reward-shaping schemes to stabilize training; applied domain randomization and dynamics perturbations to improve transfer.
• Integrated simulation policies with ROS pipelines and executed transfer runs on a low-cost robotic arm to validate real-world performance.
• Outcome: stable training recipes, transfer evaluation reports, and codebase for reuse across manipulation tasks.

• Formulated traffic-signal control as an RL problem on city-scale graphs; designed state representations and reward formulations focused on queue-length and travel time metrics.
• Implemented simulation pipelines using SUMO and Gym interfaces, trained multi-agent RL policies and analysed stability and generalization across intersections.
• Collaborated on deployment feasibility studies with industry partners (prototype controllers, latency constraints, evaluation metrics).
• Outcome: experimental results demonstrating queue reduction, reproducible SUMO+RL setups, and technical documentation for follow-up work.

• Developed deep learning pipelines for L1 identification from L2 speech, including feature extraction, augmentations and classifier design.
• Explored attention-based interpretability methods to surface salient acoustic cues and improve model explainability for linguistically-relevant predictions.
• Preprocessed and curated speech datasets, conducted ablation studies, and evaluated cross-speaker generalization.
• Outcome: conference presentation and an interpretable pipeline for L1 identification experiments.

Built data-driven document processing and automation solutions that reduced manual effort in deployments.

Optimized posture-detection pipelines and developed low-latency inference endpoints for telephysiotherapy.

Implemented skinning & rigging prototypes for 3D animation pipelines used in metaverse features.