About Me

I am a PhD Candidate in Electrical & Computer Engineering at The Ohio State University, advised by Prof. Ness B. Shroff. I work at the intersection of Machine Learning, NLP, and Systems, focusing on how to align powerful models with human preferences, keep them safe, and still run them at scale on messy, real-world infrastructure.

My path into AI started with seismic time-series during my M.S. at the University of Tehran, where I built deep models to detect and classify earthquake events in noisy waveforms. That project convinced me that real data is never clean or stationary, and that perspective has shaped how I think about every model since.

At Ohio State, I moved into Safe Reinforcement Learning with instantaneous constraints. I helped evaluate an algorithm (ICML 2025) showing that even in non-convex settings it is possible to have both exploration and rigorous guarantees about what the agent will not do. That work made me think constantly in tradeoffs: performance versus risk, optimism versus caution, and theory versus real-world execution.

More recently, I’ve been asking similar questions in the world of foundation models. In collaboration with researchers at Google, I’ve been working on COMPASS, a preference-conditioned RL fine-tuning method that trains a single language model policy which adapts its behavior based on user preference vectors rather than one fixed notion of “good.”

In parallel, through the AI-EDGE Institute, I’ve worked on federated and distributed training for LLMs — including FIRM, a multi-objective alignment method that resolves conflicts like helpfulness vs. safety locally at the client, and FLAGON, a cross-university pre-training and fine-tuning system that uses existing heterogeneous GPU clusters behind firewalls to collaboratively train models like LLaMA-3 1B (resulting in an invited paper at IEEE/ACM ToN).

Across all of this, one idea keeps me grounded: models are only useful if we can both trust their behavior and actually deploy them in the real world.

Download Curriculum Vitae

Featured Projects

COMPASS Alignment & RL
A preference-conditioned RLxF (GRPO) method that trains a single LLM policy whose behavior can be steered by a user preference vector instead of a fixed notion of “good.”
Impact: Learns to move along the Pareto front between competing objectives and generalizes to unseen preference mixes, while surfacing reward hacking via Pareto-front analysis instead of a single score.
FIRM Federated Alignment
A federated multi-objective alignment framework where each client locally balances objectives (e.g., helpfulness vs. safety) while keeping raw data on-device.
Impact: Uses an in-client regularizer to control cross-client “disagreement drift”, cuts per-round communication to ≈1/M (for M objectives), and provides the first finite-time convergence guarantees in this setting.
FLAGON Distributed Systems
A cross-university LLM training platform that lets institutions use their existing, heterogeneous GPU clusters behind firewalls and high-latency links.
Impact: Pre-trained and fine-tuned GPT-2 and LLaMA-3 1B across sites while automatically recovering from GPU preemptions — work that led to an invited paper at IEEE/ACM ToN.

Selected Publications

View Full List on Google Scholar

Education & Experience

Ph.D. Candidate, ECE 2022 - Present
The Ohio State University
Graduate Research Assistant 2022 - Present
AI-EDGE Institute / OSU

Research on scalable LLM alignment and federated systems.

M.S. Electrical Engineering 2021 - 2022
University of Tehran

Thesis on Time-Series Modeling using Deep Neural Networks for seismic analysis.

Professional Service & Leadership

Skills

Research Keywords
LLM Alignment (RLxF) Robust Evaluation Federated Learning Scalable Systems AI Safety Multi-Objective Optimization NLP
Stack
Python C/C++ PyTorch Hugging Face AWS Docker gRPC