Jeff Liu
CS & AI / Builder / Researcher
About
I'm a computer science graduate (M.S., Tufts University) with a background in economics. I focus on AI, robotics, and building practical tools that bridge technology and real-world needs. This page collects my lightweight personal projects and research demos — mostly things I tinker with on GitHub and Hugging Face.
Featured Projects
ANIMA Framework & Smart Robot Arm
An embodied AI project pairing the ANIMA cognitive framework with a self-built smart robot arm. ANIMA explores agent reasoning and action loops for physical manipulation, while the Smart Robot Arm provides the hardware platform — integrating perception, planning, and control for real-world embodied tasks.
Smart Cleaning Robot (Noetic Version)
A ROS-based autonomous cleaning robot built on TurtleBot3 that integrates SLAM mapping, frontier exploration, and full-coverage cleaning. Features a self-designed cleaning module with a dilation-based route planner and greedy path follower, plus a custom GUI control panel and real-time voice control via the Vosk model. Developed and tested on ROS Noetic + Ubuntu 20.04.
Smart Glove: Gesture-Controlled Robotics
A Raspberry Pi Pico-based smart glove that captures hand motion and finger presses via an MPU9250 IMU and FSR402 force sensors. Recognized gestures drive downstream robotics — including controlling a robot arm in real time — with ongoing extensions into sign language recognition and IoT integration. Runs on Raspberry Pi 4B + Ubuntu 22.04 + ROS Humble.
Real-Time Weather Forecasting
A CNN-based system that generates short-term, location-specific weather forecasts. Demoed live as "Tufts Jumbo Weather Forecast" — a 24-hour forecast tailored to the Tufts University campus.
VisInject: Embedding Linguistic Commands into Images
Multimodal LLMs can be hijacked by hidden instructions embedded inside images. VisInject explores how visual prompt injection attacks work against open-source MLLMs and studies defense strategies.
Color Calibration: Light-Adaptive Color ID
A light-adaptive color calibration pipeline that predicts true colors from a photo using a calibration card. Full stack on Hugging Face: live Space, trained YOLOv8 model, and curated dataset.
Model Court: Multi-Model Verification
A multi-agent consensus framework that mitigates LLM hallucinations through cross-validation across multiple models. Useful in high-stakes domains where single-model outputs can't be trusted. Published on PyPI.
A.L.I.C.E. × Generative Agents
A 2D Westworld — five AI villagers live, scheme, and dream in a simulated medieval world. They plan their days, reflect on past events, and consolidate memories each night. Built on Stanford's Generative Agents architecture.
Education
M.S. in Computer Science
Tufts UniversityFocus on AI, Deep Learning, Reinforcement Learning, and Robotics.
Open Course Studies
UC BerkeleyEnriched my knowledge in CS and AI through open courses.
Contact
- jeffliulab.com — Lab Website
- Boston Area, MA, USA
- jeff.pang.liu@gmail.com
- linkedin.com/in/liupang
- github.com/jeffliulab
- huggingface.co/jeffliulab