Nguyen K. Phan: AI Professor of Practice

01AI Expertise & Research

Generative models and machine learning, applied to hard real-world data.

My doctoral research sits exactly where this program lives: deep learning, generative AI, and the engineering needed to make models run at scale.

01 / Generative AI

Physics-guided diffusion models

Conditional and physics-guided diffusion models for scientific data, embedding constraints like divergence-free conditions and Helmholtz-Hodge decomposition with geometric priors to produce physically plausible outputs.

Diffusion ModelsCNNsPyTorch

02 / LLM Integration

Human-in-the-loop LLM systems

Embedding large language models into 3D rendering and visualization pipelines to enable natural-language control, prompt engineering, and human-in-the-loop AI workflows, both in research and in the classroom.

LLMsPrompt EngineeringAgentic

03 / Web & Interfaces

Web-based interfaces for AI research and deployment

I build fully interactive, cross-platform browser applications that make AI research accessible and usable: from scientific visualization tools to production web services. My research deliverables run in the browser, not just on a workstation, and I extend that same philosophy to large-scale inference by end users through deployed webapps.

Web ArchitectureCross-platformEnd-user AI

02Main Publication

IEEE Visualization & Visual Analytics (VIS) 2025

Vector Field Synthesis with Sparse Streamlines Using a Diffusion Model

A conditional diffusion model that reconstructs a complete vector field from only a handful of sparse streamlines that outperforms traditional optimization methods.

It demanded hands-on proficiency with the same deep-learning foundations the MS in AI's Machine Learning & Deep Learning course is built on: CNN architectures, diffusion-based generative modeling, and training pipelines that respect physical constraints.

DiffusionGenerative AI CNNDOI 10.1109/VIS60296.2025.00065

Sparse in

A few streamlines as conditioning input

Dense out

Full physically-plausible field, synthesized

Co-authored with mentees

R. A. Morales Vargas · S. Espriella · G. Chen , two undergraduate researchers I mentored onto the paper.

03AI in Production

My AI models run in production. I deploy and operate them for tens of thousands of users.

Exactly the production-at-scale competencies this program emphasizes: reliability, cost-aware engineering, and serving AI on real infrastructure.

BGBYE

bgbye.io ↗

80K/mo

Monthly active users

Open-source removal models served

A free, open-source image background remover serving 10 different AI models (Bria, InSpyrenet, U²-Net, Tracer, BASNet, DeepLab, ORMBG, ISNet). A React front end on Cloudflare Pages, a Python/PyTorch backend with multi-GPU acceleration (CUDA & ROCm).

React 18PyTorchCUDA / ROCmCloudflare

PlotBunni

plotbunni.com ↗

10K/mo

Monthly users

Concurrent connections on one consumer GPU

An open-source AI novel-writing suite. The headline engineering story: I deployed a production LLM on a single consumer GPU (RTX 5060 Ti) serving up to 50 concurrent connections, reaching thousands of users a month on cost-aware, right-sized hardware.

LLM ServingVite + ShadCNIndexedDBCost-aware

Both are FOSS, open-sourced on GitHub for the community.

04AI in the Classroom

I already teach with AI: as tutor, creative tool, and subject.

Across COSC 1336, 2306, and 4370 I've woven AI into how students learn, build, and explore, with a deliberate focus on making instruction intuitive and accessible.

COSC 4370 · Interactive Computer Graphics

An LLM inside the 3D rendering pipeline

I designed an assignment where students control a live 3D scene through natural language: adding and removing objects, changing textures and colors, animating elements. The real lesson is prompt engineering: students rewrite the system prompt to make the LLM manipulate the world more reliably.

Live, right now. Interact with it during the talk. Connects classic graphics to Generative & Agentic AI.

3dstudio.pages.dev · COSC 4370 LLM exercise

Course sites built for intuition and access: interactive demos for every concept

For COSC 2306 and 4370 I built dedicated lecture sites with live, interactive demos so abstract methods become something students can touch. Everything stays publicly accessible, anytime.

2306.pages.dev

COSC 2306 · Data Programming: lectures and interactive data-structure demos.

cosc4370.pages.dev

COSC 4370 · Computer Graphics: lectures and interactive rendering demos.

CodeHelp · COSC 1336: guided programming help

CodeHelp guided AI help interface used in COSC 1336

COSC 1336 · Intro Programming

CodeHelp: an AI tutor that guides, never hands over answers

In my intro course I adopted CodeHelp, an AI assistant that scaffolds students toward their own solution rather than spitting out code. Students use it freely on exercises and practice exams; it lowers the anxiety of a first programming course without short-circuiting the learning.

Built by Somaia Alhazmi, a UH Ph.D. student researching AI in education. I value collaborating with peers advancing the field.

05Teaching Philosophy

Deliver rigorous material. Build the confidence to use it.

Scaffolded learning. I break complex problems into manageable pieces and model a divide-and-conquer approach students can repeat independently.
Active, real-world framing. Hands-on problem solving and in-class discussion; every data structure motivated by a concrete use case before the formal definition.
Support beyond the classroom. Extra office-hours practice, detailed email replies, partial credit, and feedback that explains how to improve, not just what was wrong.

Inclusive teaching in practice

For a student with special needs in a visualization course, I modified the assignments, volunteered as his project groupmate, and helped him prepare and deliver his final presentation.

He completed the course with new confidence. Every student deserves the adaptations they need to succeed.

06Assessment Design & Academic Integrity

How we assess shapes what students believe matters.

Fair, scalable assessment and a serious, practical approach to integrity in an age of AI-assisted work.

Auto-grading at scale

Reliable feedback for complex work

For COSC 6344 (Scientific Visualization) I helped build an auto-grading system tailored to the Trame/VTK web framework, where standard scripts fall short, giving consistent assessment and freeing time for qualitative feedback on design decisions.

Integrity by design

Canary instructions in prompts

I embed invisible, hard-to-detect "canary" instructions in assignment and exam prompts. Submissions that blindly paste an AI's response surface immediately: a low-friction signal that catches misuse without burdening honest students.

Verifiable conditions

Timestamp cross-checks

A custom card-reader app logs both entry and exit times for in-person exams, cross-checked against submission timestamps, confirming work was done under exam conditions, not before or after.

Responsible AI starts with assessment students can trust is fair.

07Mentoring & Student Research

I turn students into published researchers.

I met weekly with undergraduate and high-school researchers, building confidence, teaching methodology, and guiding real contributions. Many earned co-authorship on major publications.

Mentorship is, to me, one of the most valuable things a faculty member offers, and I'm eager to do it for the MS in AI.

Student co-authors I've mentored

Brian Kim

Ricardo A. Morales Vargas

Sebastian Espriella

George Navarro

Sunny Kim

Reshmitha Muppala

08Selected Publications

Seven peer-reviewed works: IEEE VIS, Elsevier, SIGMOD.

2025

Vector Field Synthesis with Sparse Streamlines Using Diffusion Model

Phan NK, Morales Vargas RA, Espriella S, Chen G · IEEE Visualization & Visual Analytics (VIS)

2026

Interactive Exploration of Large-scale Streamlines via a Curve Segment Neighborhood Graph

Phan NK, Kim B, Zafar A, Chen G · Computers & Graphics, Elsevier

Journal

2026

Physics-Guided Diffusion for 2D Vector Field Synthesis

Phan NK, et al. · In progress

Under review · AI

2024

Curve Segment Neighborhood-based Vector Field Exploration

Phan NK, Chen G · IEEE Visualization & Visual Analytics (VIS)

Conference

2024

Growing a FLOWER: A Diagram Unifying Flow and ER Notation for Data Science

Ordonez C, Varghese R, Phan N, Macyna W · HILDA @ SIGMOD

Workshop

2023

FCLWebVis: A Flexible Cross-Language Web-based Data Visualization Framework

Phan NK, Chen G, Navarro G, Muppala R, Chu J, Kim S · VDA, IS&T Electronic Imaging

Conference

2022

Direct Neighbor Search for Curve-based Vector Field Processing

Phan NK, Chen G · IEEE VIS & VDA, IS&T (Posters)

Poster

09What I'd Bring to the MS in AI

Ready to teach the core. Eager to build what's next.

Courses I can teach now

Machine Learning & Deep LearningDiffusion · CNN · LLM

AI Systems Engineering & Deployment80K-user proof

Generative, RAG & Agentic AICurriculum dev

Mathematical Foundations for AILinear algebra · probability

Responsible AIEthics · integrity

What I want to try

RAG-powered course assistants: a CodeHelp-style tutor grounded in each course's own lecture notes, so help is always on-syllabus.
Agentic AI capstones with Houston industry: energy, the Texas Medical Center, and NASA on our doorstep are real problems students can solve.
Production MLOps in the classroom: students ship a model to real users and own reliability and cost, the way I did with BGBYE and PlotBunni.
Cost-aware AI on right-sized hardware: my single-GPU LLM deployment becomes a live teaching case in inference optimization.

Let's build it together.

Nguyen K. Phan, Ph.D. nguyenpkk95@gmail.com +1 (503) 516-8474 · Houston, TX