Workstation AI Solution Engineer · Lenovo / MS CS · Georgia Tech
I work on computer-use agents, proactive systems, and multimodal RAG — from research benchmarks to on-device deployment on AI PCs. The part I care about most is the last mile: making intelligent systems reliable enough to ship on real hardware, for real users.
01 Flagship research
A GUI-based environment of 200 tasks across 11 mechanical CAD categories. Agents work through screenshots and executable UI actions, and are scored by execution-based checks on saved, verifiable FreeCAD artifacts — not by looks-right heuristics.
The gap is the point — CADWorld measures how far agents still are from expert spatial work.
02 Selected work
Small enough to run, structured enough to trust, useful enough to meet people inside real tasks.
GUI agents often compute the right action after the window has closed. AAPT compiles a bounded policy tree while the screen is quiet, then routes live frames with a low-token observer — moving generative decode off the decision-time critical path and winning a contested 600 ms window (0.79 vs. 0.50, pre-registered).
A cross-platform tool that captures and replays desktop UI actions across Linux, Windows, and macOS — with Python and C++ capture paths — to build behavior-cloning datasets for computer-use agents.
Memory that persists beyond the context window: SCoL turns useful context into durable model knowledge, so assistants stop forgetting what matters.
Structured knowledge, reactive retrieval, and background monitoring that surfaces suggestions before you ask.
How CS1 students perceive LLM-generated formative feedback on Java assignments — human-centered evidence for AI in education. Published at AAAI.
Multimodal retrieval-augmented generation for manufacturing QA across text, images, tables, and formulas — evaluated on 1,515 QA pairs.
03 Publications
Ten publications and preprints: AI capability paired with deployment reality.
200 GUI tasks in 11 mechanical CAD categories, scored by execution-based checks on saved FreeCAD artifacts. Best agent: 25.0%; human experts: 87.0%.
Adaptive Anticipatory Policy Trees (AAPT): a frozen multimodal model compiles bounded policy trees during quiet screen phases; a low-token observer fires pre-authorized actions inside the event window. Pre-registered, capability-gated results across six models.
04 Experience
From mechanical systems and lab research into applied AI — shaped by real users, hardware constraints, and cross-functional teams.
Building and deploying AI PC and workstation experiences: local model deployment, proactive agent workflows, and next-generation productivity systems.
Commercial LLM testing, local deployment, dataset collection, and efficient adaptation methods including QLoRA-style fine-tuning workflows.
Air and water-cooling experiments, TCL and Python analysis tooling, and testing-efficiency improvements for server thermal engineers.
AI education, generative intelligent computing, ultrasonic material characterization, biomechanics, and biomedical engineering projects.
“AI should be technically serious and personally useful. The best systems make people feel more capable, not more managed.”
Design principle — agents, AI PCs, everyday intelligent tools
05 Contact