Featured Engineering Works
Spanning governed local-first agent operating systems, high-concurrency enterprise recruitment backends, ML resume scoring engines, and high-throughput inference runtimes.
β‘ Heterogeneous MaaS Platform (K8s + GPUStack Serving Cluster)
An enterprise Model-as-a-Service (MaaS) computing substrate engineered for large-scale LLM deployment. Built on Kubernetes (K8s) and GPUStack to virtualize and schedule heterogeneous compute clusters. Deep operational mastery in multi-chip unified model serving across NVIDIA (A800 / H800), Huawei Ascend (910B / 310P NPUs), and Iluvatar (PPU 810E) accelerators; tuned vLLM and MindIE runtime kernels, weight distribution caching, dynamic batching, and multi-tenant resilient routing.
π€ Navi - Governed Agent OS
Personal open-source flagship project: A local-first, governed personal AI assistant Agent OS. Powered by declarative schema-validated capability manifests for task planning, durable Goal/Run state machines, scoped typed memory with conflict resolution, and Linux Bubblewrap (bwrap) sandbox isolation. Shipped with CLI, authenticated local FastAPI services, and native WeChat & Telegram connectors.
π’ Nowcoder Enterprise Recruitment & AI Resume Scoring
Core engineering for a commercial B2B recruitment & talent acquisition platform. Developed enterprise backend services using Java / Spring Boot microservices and high-concurrency database designs, supporting massive enterprise job postings, candidate pipelines, and complex RBAC. Designed and deployed an AI resume scoring & matching system based on Python scikit-learn (sklearn) Support Vector Machine (SVM) and multidimensional text feature engineering.
π€ Doraemon Code (Autonomous Coding Agent)
A full-featured open-source local coding agent CLI, engineered as an open alternative to Claude Code. Implements plan, review, and build execution modes with a decoupled agent runtime managing state loops, task traces, and dynamic execution. Features native SDK integrations for OpenAI, Anthropic, Google, and OpenAI-compatible upstreams (vLLM/Ollama); ships with 39+ workspace tools, prompt-activated Skills, and Model Context Protocol (MCP) remote server extensibility.
π¨ Joyarti 3D Customizer (Commercial AI Generative Platform)
Full-cycle commercial product: An e-commerce and generative platform for personalized 3D-printed figurines and custom art gifts. Architected the core AI image generation pipeline, orchestrating multi-stage workflows via ComfyUI (text-to-image, LoRA style adaptation, ControlNet pose/geometry conditioning, and facial identity preservation). Built high-concurrency Java backend task queues and GPU cluster dispatching to deliver industrial-grade assets for 3D manufacturing.
π openDataAnalysis (Autonomous Tabular Data Agent)
An open-source interactive data analysis platform powered by an autonomous agent. Moving beyond static DAG pipelines, it employs dynamic tool-calling and self-directed planning to analyze tabular datasets (CSV/Excel uploads or PostgreSQL/MySQL bounded snapshots). The agent autonomously explores schemas, executes sandboxed Python/SQL code, generates interactive ECharts visualizations, and synthesizes structured analytical reports. Built on Go, Vue 3, and isolated Docker execution containers.
β‘ LLM Systems Architecture (Static Site Engine & Architecture)
The open-source production codebase powering this bilingual AI systems engineering hub. Engineered on a custom Vite, Handlebars, and Marked build pipeline featuring native KaTeX mathematical typography and syntax highlighting. Ships with zero-overhead client-side inverted search indexing, automated dynamic SVG OpenGraph generation, Schema.org (TechArticle & FAQPage) JSON-LD injection, and automated IndexNow protocol submission in CI with strict smoke test suites.
ποΈ Edge-AI Vision System (Edge Real-time Vision Engine)
A high-performance real-time object detection and multi-target tracking system engineered for resource-constrained edge computing hardware (NVIDIA Jetson / ARM NPU). Features YOLO architecture optimization in PyTorch, TensorRT INT8 / FP16 quantization with kernel fusion, and zero-copy OpenCV hardware video stream decoding. Integrates ByteTrack multi-object tracking to deliver sub-millisecond inference and reliable edge-to-cloud telemetry synchronization.
π’ Smart Campus AI (Video Stream Intelligence & IoT Dispatch)
An enterprise end-to-end multi-channel video analytics and IoT automated dispatch system. Built with an asynchronous Python microservices architecture and distributed event buses to handle high-throughput RTSP/RTMP IP camera stream ingestion, adaptive hardware decoding, and asynchronous inference queues. Powers perimeter intrusion alerts, crowd density estimation, and anomaly detection, triggering real-time WebSocket notifications, snapshot archiving, and multi-channel incident dispatches.