Harshit Bokadia

Research Associate · SUTD  |  Canadian Citizen

I am a Research Associate at SUTD (Singapore University of Technology and Design). My research builds AI pipelines for real-time estimation of learners' internal world models from motion-capture kinematic data, modelling movement exploration as a Markov Decision Process, running Bayesian inverse RL to recover reward structure and state-transition dynamics, and using diffusion models and inferred world models to scaffold open-ended embodied creativity in real time. This work connects foundational questions in world model learning and generative modelling to the physical, embodied setting.

My background spans a deliberate sequence of research and engineering roles, each building toward my current focus:

• 2008–2014BTech, Mechanical Engineering followed by nearly two years as a Project Engineer in Industry 4.0 manufacturing in India.
• 2016–2018MS, Industrial and Systems Engineering, Rutgers — depth in optimization, statistics, and systems thinking; thesis on deep learning for semiconductor process control.
• 2018–2020Computational neuroscience and cognitive science, Rutgers — electroencephalography (EEG), motor control and sensorimotor integration, Brain-Machine Interfaces (BMI), neural connectivity analysis, and digital biomarkers for autism including the Autism Diagnostic Observation Schedule (ADOS).
• 2020–2021Explainable AI (XAI) research under DARPA (Defense Advanced Research Projects Agency) / AFRL (Air Force Research Laboratory), Rutgers-Newark — novel perceptual and semantic interpretability metrics for medical imaging.
• 2021NeuroAI Intern, Mila — neural ODE modeling, EEG analysis, and generative modeling of neural dynamics.
• 2022Medical AI group, University of Waterloo — vector-symbolic architectures for structured knowledge encoding in clinical NLP.
• 2023–2025Research Engineer, Holland Bloorview Kids Rehabilitation Hospital — clinical data science for neurodivergent populations; medication recommendation AI; biosignal processing; large-scale POND (Province of Ontario Neurodevelopmental Disorders) network clinical trial data.
• 2025Data Scientist, Waypoint Centre for Mental Health Care — production AI deployment in healthcare; EHR integration via Azure ML; regulatory-compliant ML pipelines.

My current work sits at the intersection of these threads and points toward where I want to go next: robot learning, robot foundation models, and diffusion-based generative models for embodied agents, world model architectures that generalize across tasks and embodiments, and the post-training and inference machinery that makes large models deployable and trustworthy. My interpretability work, building evaluation frameworks for explanation faithfulness under DARPA/AFRL, gives me a concrete foothold in mechanistic interpretability of LLMs: understanding what circuits and features actually encode, not just what outputs look like.

Research interests

• Embodied AI / Robotics World model learning and representation in agents · robot learning and robot foundation models · diffusion models for embodied action generation · model-based RL for generalization under partial observability · inferring latent dynamics from high-dimensional behavioral data
• LLMs / MLLMs Post-training (RLHF — Reinforcement Learning from Human Feedback; DPO — Direct Preference Optimization; instruction tuning) · inference efficiency (speculative decoding, vLLM, quantization) · VSA (vector-symbolic architecture)-structured knowledge encoding
• Interpretability Mechanistic interpretability of transformer circuits · feature geometry and superposition in LLMs · evaluation frameworks for explanation faithfulness (background: saliency metrics, semantic overlap, DARPA XAI)
• Clinical AI Biosignal processing (EEG, accelerometry) · medication recommendation in neurodivergent populations · regulatory-compliant ML deployment