research

machine unlearning for safe and adaptive robotics — supervised by prof. triantafillou, university of warwick. investigating selective forgetting mechanisms for robots operating in dynamic environments.

lévy-driven pides via physics-informed neural networks — developing convergence guarantees for neural solvers on lévy-driven partial integro-differential equations. targeting applications in stochastic volatility modelling.

rxl — efficient rl kernels library — c++20 library for reinforcement learning with mathematical complexity guarantees. cuda-accelerated kernels with pybind11 bindings. framed around provable efficiency, not empirical benchmarking.

current deep-dives:

• sde numerics: numerical schemes for lévy processes and stochastic volatility models.
• kernel optimisation: mathematical proofs of why ml systems are fast, not just empirical findings.
• hardware-aware ml: cuda kernel design and c++ systems for ultra-low latency execution.