Projects
Selected research and engineering projects across AI and systems.
Research
-
Multi-dimensional Fine-grained Reward-driven LLM Mathematical Reasoning
Project Lead · Dec 2024 – May 2025
This project tackled enhancing LLM mathematical reasoning under tight resource constraints. We developed a novel reinforcement learning approach centered on a multi-dimensional reward function, which precisely guides the model's learning process. This strategy yielded a significant 4.03% accuracy boost on GSM8K using just a single GPU.
Key Achievements
- ✓ Achieved a 4.03% accuracy gain on GSM8K over a strong baseline (8.43% total) via RL fine-tuning.
- ✓ Designed a five-component reward function family (accuracy, format, length, pattern, efficiency) for granular feedback.
- ✓ Demonstrated efficient fine-tuning on a single consumer GPU (RTX 3090) using Unsloth and QLoRA.
Unsloth QLoRA GRPO PyTorch Transformers RL -
Multimodal Risk Content Identification Technology
LLM Team Lead · Jun 2024 – Nov 2024
This project architected an award-winning system for multimodal risk content identification. It introduces a sophisticated workflow that decomposes complex audit tasks into manageable stages. By employing dynamic tree parsing and staged inference, the system significantly enhances the efficiency, granularity, and generalization of threat detection.
Key Achievements
- ✓ Awarded National 2nd Prize in the 'Challenge Cup' competition (LLM track).
- ✓ Engineered a dynamic tree parsing mechanism for fine-grained, recursive risk analysis.
- ✓ Designed a multi-stage workflow with concurrent inference and staged routing to optimize resource usage.
LLM CoT LMDeploy AWQ Dify Qwen2-VL InternVL2 -
Music Emotion Recognition Support Technology Research
Principal Investigator · Mar 2023 – Mar 2024
This research project tackled key challenges in music emotion recognition (MER) by inventing a novel annotation method. We proposed and implemented a point trajectory system on the Valence-Arousal plane, which captures dynamic emotional changes with high temporal density and granularity. A full-stack platform was built to facilitate this method.
Key Achievements
- ✓ Invented a point trajectory annotation method, improving data density (National Invention Patent, 1st Inventor).
- ✓ Designed the methodology based on the Valence-Arousal dimensional model of emotion.
- ✓ Independently developed a full-stack Music Emotion Labeling (MEL) platform using Django and jQuery.
Django LayUI jQuery MySQL COS CDN
Engineering
-
JChat: AI Super Terminal for Developers
Project Lead · May 2025 – Present
An AI productivity engine designed as a super terminal for developers. It solves critical bottlenecks in complex AI tasks through three core innovations: an ultra-long context editor powered by Monaco, a batch workflow automation engine, and a 100% local-first architecture for absolute data privacy and offline access.
Key Achievements
- ✓ Integrated Monaco Editor for a crash-free, IDE-like experience with million-token contexts.
- ✓ Designed a 'Group Session' workflow engine for batch file processing and command application.
- ✓ Engineered a 100% local-first architecture with IndexedDB for privacy, speed, and offline access.
Next.js TypeScript IndexedDB Monaco Editor React -
CLIP-based Content-Based Image Retrieval System
Project Lead · May 2024 – Jun 2024
This project delivered a high-performance, cross-modal image retrieval system for million-scale datasets. By integrating OpenAI's CLIP embeddings with an optimized vector database (pgvector), the system achieved a 160x reduction in query latency, enabling near real-time search.
Key Achievements
- ✓ Reduced Top-K retrieval latency from ~2s to 12ms by building an HNSW index for ANN search.
- ✓ Performed large-scale feature extraction on the Unsplash dataset using CLIP's zero-shot capabilities.
- ✓ Built a scalable vector database architecture using PostgreSQL with the pgvector extension.
CLIP pgvector PostgreSQL HNSW Python ANN
Systems
-
C++ Unix-like File System Implementation
Project Lead · Jun 2024 – Jul 2024
A POSIX-compatible, Ext-like file system was designed and implemented from scratch in C++. The project covers core file system concepts, including inode structures, block management, and caching, and is integrated with a modern CI/CD pipeline for automated testing and delivery.
Key Achievements
- ✓ Implemented a complete file system with superblock, bitmaps, inode table, and data blocks.
- ✓ Engineered multi-level address indexing for large file support and a Dentry cache for faster path lookups.
- ✓ Designed efficient block allocation and reclamation mechanisms to manage disk space.
C++ POSIX Filesystem CI/CD Gitea Actions -
Integrated Cloud Server Cluster Monitoring Platform
Agent Team Lead · Apr 2024 – May 2024
To improve remote operations for server clusters, this project delivered a comprehensive monitoring solution. A lightweight, zero-touch agent was developed for intelligent service discovery and robust, real-time telemetry of over 100 system metrics, ensuring high availability and fault tolerance.
Key Achievements
- ✓ Engineered a zero-touch agent deployment with PKI & JWT dual authentication.
- ✓ Enabled real-time telemetry for 100+ performance metrics via intelligent service discovery.
- ✓ Designed a resilient data pipeline using a FIFO queue, local SQLite caching, and batch reporting.
Shell Linux Crontab JWT PKI FIFO SQLite