Experience
Research and education timeline.
PhD Student
My research connects identifiability theory with mechanistic interpretability. I develop methods that interpret unsupervised representations by operationalising identifiability guarantees, and I build evaluation protocols that remain meaningful under structural misspecification. Much of this work targets large language models (LLMs), where I aim to formalise and stress-test causal claims about circuits and representations.
advised by Dhanya Sridhar
Research Intern
Adapted language models for text-based reinforcement learning environments (Jericho), where the agent must parse natural language observations, maintain belief state, and act in a combinatorially large action space. This was sequential decision-making under language before large-scale pretraining made the language part easy—model capacity was scarce and architectural choices were consequential. The experience posed the same question: what separates a model that tracks a useful latent state from one that merely correlates with low-level observable features?
Master's Student
Studied how agency shapes what is learnable: how model-based RL agents can acquire (and whethe they already do) causal world models through interaction with the environment. Built controlled physics simulations to evaluate planning and long-term memory under distribution shifts and novel interventions in partially observable, stochastic environments. This work highlighted the question of enabling interaction-guided or affordance-based representation learning, i.e., what a system can do and observe as a result of it, such that learned variables reflect this underlying structure of control?
advised by Devon Hjelm
Research Programmer
Empirical Inference, Max Planck Institute for Intelligent Systems
Developed real-time control and logging software for robotic manipulators and was the primary developer of the accompanying simulation package. Research on self-supervised representation learning, specifically systematic generalisation. This raised the question I have worked on since: how to learn and evaluate whether a representation has genuinely captured any relevant structure from data (and what relevant is).
Bernhard Schölkopf · Stefan Bauer
Bachelor's in Electrical Engineering
Indian Institute of Technology Kanpur
Research exploration across convex optimization, stochastic processes, robotic manipulation, and formal methods. Projects included building physics simulations (BulletPhysics, C++) for deformable object manipulation; formulating dynamic programs for supply-chain decisions; studying online convex optimization; working with stochastic geometry and spatial point processes; and using Z3 to formalize reasoning in multi-agent social deduction.
Dmitry Berenson (deformable object manipualtion, University of Michigan) · Amar Sapra (supply chain, IIM Bangalore) · Indranil Saha (Z3 multi-agent reasoning, IIT Kanpur)