Jeff Liu — Agentic AI Solutions

Projects项目

Solutions, demos, and technical projects showcasing end-to-end delivery. 解决方案、演示和技术项目，展示端到端交付能力。

Color Calibration Solution色彩校准解决方案

● Live Demo

Interior Design & Quality Inspection — Color recognition fails under variable lighting. Built an MLOps pipeline with multi-model comparison for robust color ID. Deployed as a real-time web tool.室内设计与质量检测 — 不同光照下颜色识别失败。构建了MLOps流水线，通过多模型对比实现精准颜色识别，已部署为实时Web工具。

MLOpsML PipelineMulti-ModelReal-Time

Live Demo GitHub

Smart Cleaning Robot: Autonomous Coverage智能清扫机器人：自主全覆盖方案

● Open Source

Service Robotics & Smart Home — Commercial robots follow random paths, leaving blind spots. Built intelligent cleaning with frontier exploration, optimal path planning, and multi-modal control.服务机器人与智能家居 — 市面清扫机器人采用随机路线，存在清扫盲区。构建了前沿探索式自主建图、最优路径规划和多模态控制的智能清扫方案。

Frontier ExplorationPath PlanningVoice ControlROS

GitHub

Model Court: Multi-Model VerificationModel Court：多模型交叉验证框架

● Open Source

AI Compliance & Reliability — Single-model outputs can be unreliable in high-stakes domains. Designed a multi-model ensemble and cross-validation framework. Published on PyPI.AI合规与可靠性 — 单一模型输出在高风险领域不可靠。设计了多模型集成与交叉验证框架，已发布至PyPI。

Multi-ModelCross-ValidationAI Compliance

GitHub PyPI

Fact Check: LLM-Enhanced HTTPS Proxy事实核查：LLM增强HTTPS代理

● Video Demo

Information Security — Web content authenticity is hard to verify. Built an HTTPS Proxy that uses LLM to intercept and flag misleading content in real-time.信息安全 — 网络内容真实性难以辨别。构建了HTTPS代理，利用LLM实时拦截并标记误导性内容。

HTTPS ProxyLLMReal-Time

GitHub Report

AgentKit: Domain-Specific AI AgentAgentKit：领域AI智能体框架

● Video Demo

Healthcare & Finance — Complex tasks require AI systems that reason, retrieve, and act. Built an extensible Agent framework with RAG, demonstrated on medical sepsis task.医疗与金融 — 复杂任务需要AI自主推理、检索和执行。构建了可扩展Agent框架，集成RAG，在医疗败血症任务上验证。

AI AgentRAGTool-augmentedPrompt Eng.

GitHub

Coming Soon即将发布

VisInject: Visual Prompt Injection Attack & DefenseVisInject：视觉提示注入攻防研究

In Progress

AI Security — Open-source model deployments face security vulnerabilities from visual prompt injection. Researching attack vectors and building defense mechanisms.AI安全 — 开源模型部署面临视觉提示注入攻击漏洞。研究攻击向量并构建防御机制。

AI SecurityPrompt InjectionLLM Defense

Coming Soon即将发布

Real-Time Weather Prediction实时天气预测系统

In Progress

Agriculture & Logistics — Traditional weather forecasts lack hyper-local, real-time granularity needed for precision agriculture, outdoor event planning, and last-mile delivery optimization. Building a CNN-based system that predicts short-term weather from satellite image sequences.农业与物流 — 传统天气预报缺乏精准农业、户外活动规划和末端配送优化所需的超本地化实时精度。构建基于CNN的短期天气预测系统，通过卫星图像序列进行推断。

CNNSatellite ImageryTime-SeriesReal-Time

DQN: Value-Based RL on BreakoutDQN：Breakout基于值的强化学习

Implemented value-based RL methods (Q-learning, DQN) on Atari Breakout with CUDA acceleration.在Atari Breakout上实现基于值的RL方法（Q-learning、DQN），使用CUDA加速。

Q-learningDQNRLCUDA

GitHub

Gesture Recognition: Smart Glove Robot Arm手势识别：智能手套控制机械臂

● Video Demo

An AI-powered Raspberry Pi Pico-based smart glove that recognizes gestures to control a robot arm.基于Raspberry Pi Pico的AI智能手套，通过手势识别控制机械臂。

EmbeddedGesture Recognition

GitHub

PPO & Actor-Critic: Walker2d LocomotionPPO与Actor-Critic：Walker2d运动控制

Implemented policy-based RL methods (PPO, Actor-Critic) on the MuJoCo Walker2d locomotion task.在MuJoCo Walker2d运动任务上实现基于策略的强化学习方法（PPO、Actor-Critic）。

MuJoCoPPOActor-CriticRL

GitHub

MQTT & Edge AI: IoT PlatformMQTT与边缘AI：IoT平台

A scalable IoT platform for real-time monitoring and control. Uses MQTT and Raspberry Pi hub for remote setup and rule-based automation.可扩展的IoT平台，用于实时监控和控制设备。使用MQTT和Raspberry Pi中控实现远程自动化。

MQTTROSReactEdge AI

PID Control: Centroid-Based Line FollowingPID控制：基于质心的巡线算法

A robot that follows a line and avoids obstacles using color-based centroid detection and PID control.使用基于颜色的质心检测和PID控制实现机器人巡线和避障。

ROS1OpenCVPID Control

GitHub

Wall-Following Algorithm: Maze Escape沿墙算法：迷宫逃脱

A maze escape robot using left wall following algorithm, built at Brandeis Robotics Lab.在Brandeis机器人实验室构建的迷宫逃脱机器人，使用左墙跟随算法。

ROS1LinuxLiDAR

GitHub

Monte Carlo ES: Optimal Blackjack Policy蒙特卡洛ES：最优21点策略

Simulated Blackjack 2 million times using Monte Carlo Exploring Starts to converge on the optimal policy.使用蒙特卡洛探索起始方法模拟21点游戏200万次，收敛至最优策略。

RLMonte Carlo ES

GitHub

ArUco Detection: Real-Time Fiducial TrackingArUco检测：实时基准标记追踪

● Video Demo

Real-time ArUco fiducial marker detection system using computer vision.使用计算机视觉的实时ArUco基准标记检测系统。

Computer VisionArUco Markers

GitHub

LLM Agent: Intelligent NPC LearningLLM智能体：NPC自主学习

● Video Demo

A prototype AI agent in a game environment where the NPC learns all skills from the player through natural interaction.一个游戏环境中的AI智能体原型，NPC通过自然交互从玩家处学习所有技能。

OpenAI APIPrompt Eng.Pygame

GitHub

A* Search: OpenStreetMap Route PlannerA*搜索：OpenStreetMap路径规划

Implemented A* route planning on OpenStreetMap data — from Tufts University to Mystic River.在OpenStreetMap数据上实现A*路径规划——从塔夫茨大学到Mystic River。

OSMC++A*

GitHub

DS&A Roadmap数据结构与算法路线图

A structured study roadmap covering foundational data structures and algorithms, plus competitive programming links.结构化学习路线图，涵盖基础数据结构与算法以及竞赛编程链接。

Data StructuresAlgorithms

View

Coming Soon即将发布

Deep RL: Escape the Maze V2深度RL：迷宫逃脱 V2

In Progress

Training an RL agent to explore and escape unknown mazes through trial-and-error, replacing hand-crafted wall-following rules with learned policies.训练RL智能体通过试错探索未知迷宫并逃脱，用学习到的策略替代手工编写的沿墙规则。

Reinforcement LearningDeep Q-NetworkExploration

Ichiban Lottery天元一番赏

● Commercial

A WeChat Mini Program for anime figure lottery, deployed in a real market with real users, generating over $15,000 in revenue. Demonstrates end-to-end product delivery.微信小程序动漫手办抽奖平台，已在真实市场部署，累计营收超过$15,000。展示了端到端产品交付能力。

WeChat Mini ProgramProductE-commerce

Details

Technical Insights技术洞察

Technical deep dives demonstrating domain knowledge and analytical depth. 展示领域知识和分析深度的技术研究报告。

When Does Few-Shot Learning Work? A Practical Evaluation of MAML小样本学习何时有效？MAML的实践评估

Evaluates meta-learning (MAML) on Mini-ImageNet against standard fine-tuning baselines.在Mini-ImageNet上评估元学习（MAML）与标准微调基线的对比。

2025-12-15

Choosing the Right Training Strategy: Normalization vs. Regularization for Transformers选择正确的训练策略：Transformer中归一化与正则化的权衡

Systematically compares normalization and regularization techniques on Transformer models for machine translation.系统比较Transformer模型在机器翻译中的归一化和正则化技术。

2025-11-19

How Initialization and Learning Rate Scheduling Affect Model Convergence初始化和学习率调度如何影响模型收敛

Investigates how different initialization strategies and learning rate schedules impact deep learning training.研究不同初始化策略和学习率调度对深度学习训练的影响。

2025-10-27

Optimizer Selection Guide: SGD, AdaGrad, RMSProp, and Adam Compared优化器选择指南：SGD、AdaGrad、RMSProp与Adam对比

Benchmarks four major optimizers on neural network training, analyzing convergence speed and final performance.在神经网络训练中基准测试四种主流优化器，分析收敛速度和最终性能。

2025-10-09

Applying Value-Based RL to Visual Game Environments: Q-Learning and DQN将基于值的RL应用于视觉游戏环境：Q-Learning与DQN实践

Implements Q-learning and DQN on Atari Breakout with PyTorch, documenting practical challenges in visual RL.使用PyTorch在Atari Breakout上实现Q-learning和DQN，记录视觉RL中的实践挑战。

2025-05-06

Building Domain-Specific AI Agents: Architecture and Lessons from a Medical Task构建领域AI智能体：医疗任务中的架构设计与经验

Discusses the design of an AI Agent for medical sepsis, covering architecture and tool integration lessons.讨论面向医疗败血症任务的AI Agent设计，涵盖架构决策和工具集成经验。

2025-05-02

End-to-End Color Calibration: From ML Model Selection to Real-Time Deployment端到端色彩校准：从模型选择到实时部署

Documents the full journey from ML model development to deploying a real-time color calibration system.记录从ML模型开发到部署实时色彩校准系统的完整过程。

2025-03-14

Centroid Based Line Following基于质心的巡线方法

Extracting colors in HSV format and marking the centroid to achieve single color line following.提取HSV格式颜色并标记质心，实现单色巡线。

2024-11-15

A Simple Maze Escape Solution简易迷宫逃脱方案

A brief study on maze escape using wall-following algorithms.关于使用沿墙算法实现迷宫逃脱的简要研究。

2024-10-27

Pang (Jeff) Liu 刘庞（刘杰夫）

About Me关于我

Projects项目

Color Calibration Solution色彩校准解决方案

Smart Cleaning Robot: Autonomous Coverage智能清扫机器人：自主全覆盖方案

Model Court: Multi-Model VerificationModel Court：多模型交叉验证框架

Fact Check: LLM-Enhanced HTTPS Proxy事实核查：LLM增强HTTPS代理

AgentKit: Domain-Specific AI AgentAgentKit：领域AI智能体框架

VisInject: Visual Prompt Injection Attack & DefenseVisInject：视觉提示注入攻防研究

Real-Time Weather Prediction实时天气预测系统

DQN: Value-Based RL on BreakoutDQN：Breakout基于值的强化学习

Gesture Recognition: Smart Glove Robot Arm手势识别：智能手套控制机械臂

PPO & Actor-Critic: Walker2d LocomotionPPO与Actor-Critic：Walker2d运动控制

MQTT & Edge AI: IoT PlatformMQTT与边缘AI：IoT平台

PID Control: Centroid-Based Line FollowingPID控制：基于质心的巡线算法

Wall-Following Algorithm: Maze Escape沿墙算法：迷宫逃脱

Monte Carlo ES: Optimal Blackjack Policy蒙特卡洛ES：最优21点策略

ArUco Detection: Real-Time Fiducial TrackingArUco检测：实时基准标记追踪

LLM Agent: Intelligent NPC LearningLLM智能体：NPC自主学习

A* Search: OpenStreetMap Route PlannerA*搜索：OpenStreetMap路径规划

DS&A Roadmap数据结构与算法路线图

Deep RL: Escape the Maze V2深度RL：迷宫逃脱 V2

Ichiban Lottery天元一番赏

Technical Insights技术洞察

When Does Few-Shot Learning Work? A Practical Evaluation of MAML小样本学习何时有效？MAML的实践评估

Choosing the Right Training Strategy: Normalization vs. Regularization for Transformers选择正确的训练策略：Transformer中归一化与正则化的权衡

How Initialization and Learning Rate Scheduling Affect Model Convergence初始化和学习率调度如何影响模型收敛

Optimizer Selection Guide: SGD, AdaGrad, RMSProp, and Adam Compared优化器选择指南：SGD、AdaGrad、RMSProp与Adam对比

Applying Value-Based RL to Visual Game Environments: Q-Learning and DQN将基于值的RL应用于视觉游戏环境：Q-Learning与DQN实践

Building Domain-Specific AI Agents: Architecture and Lessons from a Medical Task构建领域AI智能体：医疗任务中的架构设计与经验

End-to-End Color Calibration: From ML Model Selection to Real-Time Deployment端到端色彩校准：从模型选择到实时部署

Centroid Based Line Following基于质心的巡线方法

A Simple Maze Escape Solution简易迷宫逃脱方案

Work Experience工作经历

ROS DeveloperROS 开发工程师

Computer Science Lecturer计算机科学讲师

Education教育背景

M.S. in Computer Science计算机科学硕士

M.A. in Economics经济学硕士

Open Course Studies公开课学习

Skills & Toolset技能与工具

Contact联系