Hi there! I'm a final-year Electrical Engineering student at Sharif University of Technology. My research interests include reinforcement learning, which I grew fond of during my Deep RL course, and estimation theory, particularly designing better estimators and analyzing their properties.
I'm currently at the Machine Learning Lab (MLL), working on meta reinforcement learning under the supervision of Dr. Soleymani Baghshah. My goal is to help agents adapt to new environments faster and more efficiently.
Previously, I interned at the DSP Lab at CUHK in Hong Kong, supervised by Dr. Kong.
PS: The background is inspired by Conway's Game of Life.
Started research at MLL Lab on meta-reinforcement learning for fast adaptation under non-stationary dynamics.
Completed a research internship at CUHK DSP Lab on transfer learning for EEG anomaly detection.
Joined Sharif University of Technology as a B.Sc. student in Electrical Engineering.
Research with Dr. Soleymani Baghshah at Sharif’s MLL Lab on adapting agents to unseen tasks from limited interaction. Read the METTA paper on arXiv.
Worked with Dr. Qiuqiang Kong at the CUHK DSP Lab on using Pre-trained Audio Neural Networks for EEG anomaly detection.
Meta-reinforcement learning codebase for tracking task non-stationarity.
Setup HumanUp code that was mainly published for Isaac in UniLab setting, which is a CPU simulator for robotics
My Homework solutions and project for the DGM course at Sharif
Deep generative modeling for adversarial prompt synthesis and LLM robustness.
Used deep generative models to synthesize adversarial prompts that probe robustness and failure modes of large language models. The project studied how generated prompts behave under different safety filters and what kinds of mitigation strategies can improve robustness.
The work combined PyTorch, Hugging Face Transformers, and experiment-driven evaluation.
Homework solutions and project for the Deep RL course at Sharif
Conjugate-gradient methods for convex optimization from scratch.
Implemented conjugate-gradient and Hessian-free optimization methods from scratch, targeting numerical accuracy and practical speedups over standard solver baselines.
The experiments reached optimality gaps at or below 1e-6 and were applied to portfolio optimization and sparse regression tasks.