RL for Foundation Models
Training large models with GRPO and reward modeling to reason, decide, and self-verify through outcome-only signals.
Gainesville, FL · University of Florida
AI researcher focused on intelligent systems that can perceive, retrieve, and act across multiple modalities — reinforcement learning for foundation models, agentic AI with tool-use and self-verification, vision-language models, and computer vision.
I'm a graduate researcher in Electrical & Computer Engineering at the University of Florida, working at the intersection of reinforcement learning, multimodal learning, and agentic AI. My goal is to develop AI that can reliably reason and generalize in complex, real-world environments.
I currently train large vision-language models that learn when and how to retrieve external knowledge and self-verify their answers, and I build agentic pipelines that connect perception, planning, and human-in-the-loop reasoning. Earlier work spans spatial machine learning, computer vision, and full-stack development.
University of Florida
Multimodal Learning · Agentic AI · RL · Computer VisionUniversity of Iran · GPA 4.0 / 4.0
Computer Vision · Neural Network OptimizationUniversity of Semnan · GPA 4.0 / 4.0
Web Development · Network EngineeringTraining large models with GRPO and reward modeling to reason, decide, and self-verify through outcome-only signals.
Goal-directed pipelines with tool-use, planning, retrieval, and human-in-the-loop interaction.
Multimodal systems that perceive images and ground answers in external knowledge for knowledge-intensive VQA.
Dense retrieval over tens of millions of passages, integrated with RL-trained reasoning.
Spatially-constrained self-organizing maps for real-world hotspot detection and planning.
Object detection and image understanding for vulnerability assessment from satellite & street-view imagery.
University of Florida
University of Florida
Alfred University
Pardis Tehran Science & Technology Park
A novel spatially-constrained self-organizing map with representative initialization, adaptive geographic filtering, and region-growing refinement. Applied to flood-induced E. coli contamination in Jacksonville, it detects hotspots more accurately than census tracts and standard units. Published in Scientific Reports (Nature).
A vision-language model trained with GRPO to reason over images, decide when to retrieve external knowledge, and self-verify answers. Multi-turn agentic pipeline with FAISS dense retrieval over 21M Wikipedia passages, on 4× B200 GPUs (FSDP · vLLM · Ray).
An agentic framework integrating LLM-based reasoning, planning, and human-in-the-loop interaction to generate demand-oriented regionalization for disaster adaptation — built on top of RepSC-SOM with natural-language constraint exploration.
An object detection system that identifies and classifies flood-prone architectural features from satellite & street-view imagery for campus vulnerability assessment.
Deep models predicting 42 psychological attributes from facial images with a novel activation function, plus a multimodal sentiment system combining facial analysis with Twitter text.
Explore the work behind my RepSC-SOM regionalization research two ways — a quick in-browser demo of the core Self-Organizing Map idea, or the full Regionalization Studio app I built, running live.
Pure vanilla JavaScript & Canvas — no libraries. The algorithm runs entirely in your browser.
The real Streamlit app I built: upload a CSV, map your columns, choose parameters, and run the actual RepSC-SOM pipeline — baselines, embeddings, and region growing included.
The live app runs on Streamlit's free tier — if it has gone to sleep, it may take ~30 seconds to wake up. If the frame stays blank, open it in a new tab ↗.
A recorded walkthrough of the RepSC-SOM regionalization pipeline — from data upload through region-growing refinement and hotspot detection.
I'm always open to research collaborations, PhD opportunities, and good conversations about reasoning-capable, multimodal AI. Whether you have a project in mind or just want to talk shop — I'd love to hear from you, and I usually reply within a day.