I'm a Ph.D. candidate in Computer Science at The University of Texas at Arlington, researching Agentic AI for indoor navigation, building an Agentic AI navigation coach that perceives, reasons about, and guides users through indoor environments where GPS fails. My broader work spans LLM-driven agent architectures, graph neural networks for human activity understanding, and taking models from paper to deployed production systems.

Research interests: Agentic AI • Indoor Navigation & Spatial Reasoning • LLM Agent Architectures • Graph Neural Networks • Applied ML

🔗 ayonroy.com • LinkedIn • Google Scholar • 📫 royshouhag@gmail.com

Agentic AI Navigation Coach for Indoor Navigation (Ph.D. dissertation research, UT Arlington) Designing an LLM-driven agentic system that acts as a navigation coach for indoor environments — where GPS is unavailable and users must rely on spatial understanding, landmarks, and turn-by-turn reasoning. The agent perceives indoor context, plans routes, and delivers adaptive, human-like coaching rather than raw directions. → AgenticCoach

Cognitive Fatigue Detection from Gait (UT Arlington) Developed a Residual Graph Convolutional Network (ResGCN) that detects cognitive fatigue levels from gait-cycle data, modeling skeletal walking patterns as spatio-temporal graphs.

LLM Evaluation & Comparative Analysis Published a peer-reviewed comparative study of large language models (ChatGPT versus Bard, Engineering Reports 2024), covering reasoning, factuality, and task performance.

Automated Model Selection Built and published CurFi, an automated curve-fitting tool that selects the best regression model for a dataset — open source and paired with its paper. → curfi.js

• ChatGPT versus Bard: A comparative study. Engineering Reports, vol. 6, no. 11, e12890, 2024
• CurFi: An automated tool to find the best regression analysis model using curve fitting. Engineering Reports, vol. 4, no. 12, 2022
• A Deep Learning Approach to Predict Academic Result and Recommend Study Plan for Improving Student's Academic Performance. Ubiquitous Intelligent Systems, 2021 Full list on Google Scholar

Agentic Document Intelligence at Scale — End-to-end LLM + ML system for multi-output classification: LLM clause extraction over 1,292 OCR-parsed contracts + CatBoost ensemble trained on 4M+ records, served via Databricks/MLflow and scoring 62K+ line items in production. (Trinity Industries)

Ensemble Forecasting — 4-model ensemble with 60+ engineered features, KMeans segmentation (automated k via knee detection), and per-cluster random forests producing confidence-bounded daily forecasts. (Trinity Industries)

Agentic AI & Developer Tooling — Claude Code workflows (skills, sub-agents, hooks, MCP), agent architectures with capability-based security, and VS Code tooling. → claude-config

LLM / GenAI

ML / Data

Languages

Full Stack

Cloud & Tools