Evolution had a few billion years to figure out perception, reasoning, language, and the rest of it. I’d like to skip the line. My research puts large language and vision models in the closed loop of real robots: planning, replanning when the world changes underneath them, and grounding instructions in what the robot actually sees.
First author (with B.-C. Min) · Under review · 2026
Offline audits that locate belief-state failures in probe-then-commit manipulation pipelines, before the robot commits to an irreversible action.
First author (with J. Nam, W. Jo, B.-C. Min) · Under review · 2026
LM-compiled dependencies keep cached semantic judgments valid as scenes change, so robots re-evaluate only what the change actually touches. Demonstrated on a Unitree G1 humanoid.
First author · Preprint · arXiv:2508.14006
A large-scale vector–graph corpus of 17,000 residential floor plans with rooms, walls, doors, windows, and the room-connectivity graph.
First author · Preprint · arXiv:2512.16275
A graph-conditioned generative model for functional 2D layouts that satisfies user-supplied topological constraints by construction.
First author (with A. Shah, E. Garyfallidis) · MICCAI 2026 · arXiv:2604.00250
Couples a physics-based forward model with gradient-based optimisation to recover latent fibre structure from noisy diffusion-MRI measurements.
First author (with E. Garyfallidis) · Preprint · arXiv:2602.20063
A derivative-augmented cubemap that reconstructs any function on S² with bicubic-quality from a 2×2 texel footprint, with analytic surface normals.
Core contributor & maintainer · FURY · Python Software Foundation
Core contributor and maintainer of FURY v2, with 54+ merged PRs across the WebGPU rendering core, actor primitives, and animation system.
Contributor · DIPY
Contributor: Skyline ODF visualisation in DIPY Horizon (Spherical Hermite Maps) and compute-shader streamtube generation for tractography.
The viewer renders straight from the same .pkl released
with the dataset. Toggle the typed edges (via_door,
via_window, direct, adjacency) to inspect
the connectivity graph that conditions GFLAN.
Co-designed the curriculum; taught 100+ students time-series, Fourier, audio representations, graph theory, and the NumPy / data-engineering stack.
Led labs and office hours on search, probabilistic reasoning, and ML fundamentals.
Mentored a contributor on integrating LLM-powered assistants with Discord, GitHub, and X for an open-source visualisation project.
Built play/pause/rewind, SLERP rotation, linear and non-linear interpolation, hierarchical transforms, and motion paths.
I am a PhD student in Intelligent Systems Engineering at Indiana University’s Luddy School, working in the SMART Lab with Prof. Byung-Cheol Min (previously the Garyfallidis Research Group, 2024–2026). My research is robot learning and embodied AI: LLM/VLM-based planning and closed-loop replanning, belief auditing for one-shot manipulation (Audit Before You Commit), instruction-conditioned grounding (MIND, demoed on a Unitree G1 humanoid), and generative environments for embodied agents (ResPlan, GFLAN). Earlier work includes differentiable analysis-by-synthesis (PRISM, MICCAI 2026) and real-time spherical primitives (Spherical Hermite Maps).
I co-instruct DSCI-D 321 (Data Representation), TA Intelligent Systems I, and mentor for Google Summer of Code (2024 mentor, 2022 contributor). I am a core maintainer of FURY v2 (54+ merged PRs) and a contributor to DIPY, where I shipped the Skyline ODF renderer and a compute-shader streamtube generator for tractography. I hold a BEng in Electrical Engineering (Computer Engineering & Automatic Control) from Tanta University, Egypt.
Seeking a Summer 2027 research internship in robot learning and embodied AI, and always open to collaborations and research conversations. The fastest line is email.