Developing
Multi-Agent Business Simulation
Developing simulation environments for studying interactions among autonomous business agents, integrating operational decisions, financial constraints, and long-term organizational outcomes.
PhD researcher in Computer Science, Toronto Metropolitan University.
GitHub · Google Scholar · ORCID · Email
I study how AI systems turn uncertain information into reliable decisions over time: how they acquire information, interpret it, and convert it into sequential actions — and how we can tell whether each part actually improves outcomes.
Open to collaboration. I’m interested in remote collaborations with PhD researchers working on information retrieval, agentic AI, sequential decision-making, and ML/optimization — especially projects connecting information quality to downstream decision value. Get in touch.
Developing
Developing simulation environments for studying interactions among autonomous business agents, integrating operational decisions, financial constraints, and long-term organizational outcomes.
Developing
Investigating lightweight forecasting architectures, computationally efficient learning, and generalization across diverse time-series domains.
Developing
Studying how planning, memory, retrieval, verification, and tool-use mechanisms affect the effectiveness, efficiency, and adaptability of language-model-based agents.
Developing
Exploring transferable decision-making models that can adapt across operational problems, environments, and optimization objectives.
Developing
Investigating the reliability of forecasting models under distribution shifts, with emphasis on reproducible evaluation, model selection, and out-of-distribution performance.
Conference submitted
Web agents compress long pages to fit limited context windows, but compression is judged by what it retains, not by whether the agent can still choose correctly. I study decision-valid evaluation of context: paired page states differing in one decision-critical fact, where a context passes only if the agent acts correctly on both.
Conference submitted
Retrieval and forecasting systems are usually judged by prediction quality, but an agent consumes information in order to act. I study when better-ranked or additional information actually improves downstream decisions — and when it can make them worse.
Conference submitted
Fixed pipelines — one retriever, one forecaster, one policy — cannot stay optimal everywhere. I investigate when a system should retrieve, ignore, reinterpret, or switch information sources, and when it should change the decision policy itself.
gym-invmgmt: An Open Benchmarking Framework for Inventory Management Methods.
Reza Barati, Qinmin Vivian Hu.
Accepted at WI-IAT 2026, the IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology.
Paper · Code · Package
gym-invmgmt — open benchmarking suite for inventory management: Gymnasium-compatible environments with reference classical and reinforcement-learning policies. GitHub · PyPI · Trained weights
PhD Research · 2024–Present
Developing intelligent agents that integrate information retrieval, forecasting, reinforcement learning, and operations research to make adaptive, reliable sequential decisions under uncertainty.
MSc Thesis · 2024
Research on efficient foundation models for time-series forecasting, emphasizing zero-shot generalization and computational efficiency.
Reviewer for Engineering Applications of Artificial Intelligence (Elsevier, 2 reviews in 2026)
Reza Barati is a PhD researcher in Computer Science at Toronto Metropolitan University, working on sequential decision-making under uncertainty. He is supervised by Dr. Qinmin (Vivian) Hu (Google Scholar).