Aug 2026 – May 2028
Carnegie Mellon University
M.S. in Information Networking
Graduate studies
A LITTLE CONTEXT
About
I’m pursuing an M.S. in Information Networking at Carnegie Mellon University, after graduating from UIUC with Highest Distinction in Mathematics and Computer Science. My work connects AI agents, distributed systems, and end-to-end product engineering.
Download my full résuméA foundation
Where I’ve
A foundation
for what’s next.
Aug 2022 – May 2026
University of Illinois Urbana-Champaign
B.S. in Mathematics and Computer Science
GPA 3.88 / 4.0 · Highest Distinction
Selected coursework: University of Illinois Urbana-Champaign
Data Structures · Linear Algebra · Computer Systems · Algorithms · Numerical Methods · Machine Learning · Database Systems · Machine Learning Systems · Distributed Systems · AI Agent · Compilers
Where I’ve
been building.
Jun 2026 – Aug 2026Nanjing, China
TetraBot
R&D Intern
Rebuilt an AI fire-response platform around a deterministic runtime, from agent orchestration to edge vision and a live 3D interface.
66%less incident-handling time
Read the full story: TetraBot
- Re-architected a legacy LLM-agent-orchestrated fire-response system into a deterministic state-machine kernel with deploy-time AI agents in Python, reducing incident handling from 497.5s to 167.9s (66%).
- Built a runtime on Python asyncio, Redis Streams, and SQLite: first-response latency fell from 19s to 36ms, with 3.7× throughput under concurrent alarms, zero-loss kill -9 crash recovery, and SLA-timeout red lines.
- Implemented a lightweight agent runtime with bounded tool-calling loops, JSON Schema validation, per-persona permission whitelists, resumable on-disk sessions, and audit logging.
- Designed a workflow DSL and LLM compiler to write building emergency-response plans, then verify them against safety rules and simulated fire drills in a generate–verify–repair loop.
- Built a conversational ingestion agent for device sheets, floor-plan images, and CAD drawings, reducing onboarding from days of scripting to one human-confirmed dialogue session. Shipped a voice copilot and a React + Three.js building view with live fire localization.
- Deployed YOLOv8n fire/smoke detection on the RK3588 NPU using ONNX → RKNN INT8 quantization, increasing inference from 2 to 41 FPS at FP16-equivalent accuracy and serving detections over HTTP into a multi-frame analysis chain.
Jul 2025 – Sep 2025Hangzhou, China
Machine Intelligence Lab · Westlake University
Research Intern
Explored offline reinforcement learning for long-horizon, high-precision robotic assembly on FurnitureBench.
Offline RLfor robotic manipulation
Read the full story: Machine Intelligence Lab · Westlake University
- Researched offline reinforcement learning for long-horizon, high-precision robotic manipulation on FurnitureBench furniture-assembly tasks.
- Built an end-to-end training and evaluation pipeline with Python, PyTorch, and IsaacGym, adapting the Reinformer sequence-modeling architecture and fusing multimodal inputs for high-dimensional control.
- Implemented Max-Return Sequence Modeling with expectile regression to address trajectory stitching failures in standard Decision Transformers; stabilized optimization with entropy-temperature tuning and gradient clipping.
Oct 2024 – Dec 2024Champaign, IL
National Center for Supercomputing Applications
Full-Stack Software Engineer
Built EACUE, a mobile-first platform that makes physical accessibility evaluations easier to collect and manage.
EACUEaccessibility evaluation platform
Read the full story: National Center for Supercomputing Applications
- Built EACUE, a mobile-first survey platform for recording accessibility evaluations across physical venues, replacing fragmented manual collection with a centralized digital system.
- Engineered a Next.js, Node.js, and MongoDB architecture for end-to-end survey delivery, structured storage, and CRUD workflows for users, venues, and evaluation records.
- Designed a dynamic question-routing engine that adapts survey flow to real-time user selections, improving form efficiency and reducing unnecessary input steps.
THE TOOLKIT
Different tools. Same care.
GOOD WORK STARTS WITH A CONVERSATION.