Hojat Salehi

In the decade before my Ph.D., I worked as a research fellow developing second-order gradient-based optimization methods for control problems in hybrid dynamical systems and legged robots, then built an ML pipeline for stock-market time-series prediction using LSTMs, with a stretch of industrial control engineering in between. That decade gave me both the theoretical grounding and the hands-on implementation skill I’ve carried into every project since: proving a result on paper, then shipping the system that has to hold up in practice.

My current research focuses on four areas, always pairing a theoretical result with a working system:

• Agentic AI & System Identification: multi-agent LLM systems — including retrieval-augmented (RAG) pipelines — for real-time decision-making and physical plant identification (e.g., Agentic-SysID, Game Analyst).
• Explainable Graph Machine Learning: addressing structural distribution shifts and creating explanation-preserving augmentations for semi-supervised learning.
• Privacy-Preserving Distributed Learning: applying information theory to build differentially private federated learning mechanisms using common randomness (e.g., CorBin-FL).
• Generative Models for Structured Data: mathematically grounded approaches to synthetic data generation.

I am currently authorized to work in the US through OPT (eligible for STEM extension) and require no employer sponsorship.

Three systems, with real figures and results:

CorBin-FL

Differentially private federated learning, one bit per weight.

View System GitHub

Agentic-SysID

A multi-agent pipeline that identifies control-ready models under an experiment budget.

View System GitHub

Game Analyst

Adversarial multi-agent debate for NBA/MLB game analysis, graded against real outcomes.

View System GitHub