Projects

A per-site Chrome extension that blocks search-box distractions, including placeholder text, suggestion dropdowns, and browser autocomplete.

• Chrome Extension
• JavaScript
• Productivity
• Browser UX

• Works on a per-site basis instead of forcing a global browsing rule
• Removes multiple layers of search-box distraction at once
• Keeps the browsing experience minimal and focused

A tool that lets you start a local or Hugging Face model instantly and use it directly from chat through an OpenAI-compatible endpoint.

• Local LLM
• vLLM
• OpenAI-Compatible API
• Model Serving

• Starts local or Hugging Face models without leaving the chat workflow
• Exposes the model through an OpenAI-compatible API endpoint
• Supports quick status checks, chat interaction, and server control commands

A Chrome extension that helps users speak freely while keeping their messages gentle, empathetic, and constructive through AI-generated high-EQ replies.

• Chrome Extension
• AI Assistant
• Communication
• UX

• Generates high-EQ replies based on relationship, tone, history, and intent
• Works as a floating assistant across websites such as WeChat Web, Slack, and Gmail
• Supports one-click copy for quick insertion into ongoing conversations

An efficient collaborative training framework for large language models, designed to support scalable pretraining and fine-tuning with a user-friendly workflow.

• PyTorch
• DeepSpeed
• 3D Parallelism
• PEFT

• Supports data, tensor, and pipeline parallelism for large-scale LLM training
• Integrates memory-efficient fine-tuning methods such as PEFT and LOMO
• Keeps a HuggingFace-like interface while simplifying distributed training setup

A lightweight local evaluation tool for large language models that supports side-by-side chatting, debugging, and response quality assessment across multiple models.

• LLM Evaluation
• Python
• Local Deployment
• UI

• Enables interactive comparison across multiple chat models in one interface
• Supports debugging, annotation, vote analysis, and session export
• Can be deployed locally to preserve chat data privacy

A multimodal safety reasoning project that advances safety and general capability together through safety-oriented reinforcement learning and verifier-guided post-training.

• Multimodal
• Safety Alignment
• Reinforcement Learning
• Verifiers

• Built with the SafeLadder framework for progressive safety-oriented RL post-training
• Develops intrinsic safety reasoning and self-reflection instead of only learning human preference signals
• Improves safety benchmarks strongly while preserving general multimodal capabilities

An open-source red teaming framework for multimodal large language models, built for attack generation, safety evaluation, and configuration-driven experimentation.

• Multimodal
• Red Teaming
• Safety Evaluation
• Python

• Supports 42+ black-box and white-box attack methods for MLLMs
• Uses a modular plugin architecture for attacks, models, judges, and evaluators
• Runs reproducible experiments through YAML-based configuration files