PhD Candidate · Computer Science · NTU Singapore
Trustworthy RL · Neurosymbolic AI
I’m a PhD candidate in Computer Science at NTU Singapore, advised by Arvind Easwaran (NTU) and Blaise Genest (CNRS IPAL, France), working on trustworthy reinforcement learning for safety-critical systems: making RL policies transparent, safe and verifiable. In particular, I work on neurosymbolic techniques, especially symbolic policies that are amenable to formal verification, and on learning methods to train them.
More broadly, I’m interested in neurosymbolic AI that integrates intuitive (System 1) learning with explicit, deliberative (System 2) reasoning, and in the maths underneath it all: learning theory and statistics.
Before NTU, I did an M.Tech in Computer Science at ISI Kolkata and a B.Tech in Electronics & Instrumentation at NIT Silchar.
An exact, approximation-free encoding of oblique decision trees as neural networks, so trees train end-to-end with vanilla gradient descent. The resulting tree policies are interpretable and amenable to formal safety verification.
Neurosymbolic hierarchical RL that builds its abstract world model from incrementally added knowledge, rather than learning it upfront, improving sample efficiency on sparse-reward, long-horizon tasks.
I’ll submit my thesis around Dec 2026 and am looking for roles from Feb 2027 in neurosymbolic RL for trustworthy deployment: differentiable symbolic methods for interpretability and safety, and symbolic knowledge for sample efficiency, with applications in robotics and energy systems.
From bandits and MDPs to modern deep RL, with from-scratch code for every algorithm, written to be read alongside the maths. Work in progress; code at deep-rl-jutsu.
Released RLFT — notebooks training nanoGPT on digit arithmetic, then GRPO fine-tuning for subtraction.
Neurosymbolic Reasoning with Incremental Knowledge for Sample Efficient Hierarchical RL accepted at ECML-PKDD 2026. We learn the abstract world model incrementally and propose Belief World Tree Search (BWTS).
Approximation-Free Differentiable Oblique Decision Trees accepted at JMLR (vol. 27) — extends DTSemNet to regression and continuous-action RL.
Built an F1Tenth ROS 2 race stack, up to pure pursuit.
Appointed Vice President, Graduate Students’ Club (CCDS), NTU.
Programmatic RL for Trustworthy Microgrid Management at the AI4X workshop.
DTSemNet — training oblique decision trees with vanilla gradient descent — accepted at ECAI 2024.
Co-organized NTU’s first Deep Learning Bootcamp.
Appointed Director of Career and Development, Graduate Students’ Club (SCSE), NTU.
New post on Interpretable Reinforcement Learning.
Started my PhD at NTU Singapore with Arvind Easwaran and Blaise Genest, funded by the DesCartes programme.
