Projects
Projects are grouped by direction below. Private repositories are listed by name only, without links.
Medical imaging AI
Breast ultrasound classification (bucad)
Binary benign/malignant classification on breast ultrasound images. Trained on BUS-BRA and tested externally on BUSI, producing lesion region hints, risk stratification and Grad-CAM visualisation for review.
Responsibilities: implementing and tuning the CNN classifier, fixing the evaluation metrics and threshold policy, building the Grad-CAM module, and packaging the inference demo. This work supported three competition awards.
Stack: Python · PyTorch · medical imaging · Grad-CAM
Repository
Pulmonary nodule classification and interpretable risk grading (ct-nodule-risk)
A provincial undergraduate innovation training project: classify pulmonary nodules on chest CT, grade malignancy risk, and give interpretable evidence for the decision.
I handle data processing and model experiments — nodule extraction and preprocessing, the training and evaluation pipeline, and keeping records reproducible (fixed splits, seeds and evaluation scripts). Ongoing; the current focus is making the evaluation protocol solid.
Stack: Python · PyTorch · 3D medical imaging · interpretability
Repository
Osteomyelitis fluorescence-guided diagnostic workstation (osteo-vision)
A physician workstation for fluorescence-guided osteomyelitis imaging: image review, lesion annotation and 3D display. Frontend is a Vue3 workstation UI, backend is FastAPI, and the 3D view runs on Three.js.
Because the competition required delivery on a disc, there was plenty of engineering beyond the algorithms: two build configurations (GPU accelerated and CPU fallback), cross-machine dependency and path checks, and automated regression over the whole interface (13 pages, 265 buttons, 192 actual clicks, zero failures).
Stack: Vue3 · FastAPI · Three.js · Electron · Playwright
Repository
SWUST Code — a terminal-native coding agent (swust-code)
A personal build based on an open-source coding agent, extended with persistent memory, multi-agent workflows on smaller models, and goal-driven autonomy, plus its own documentation site. It is usable day to day and has been used to develop other projects.
Stack: TypeScript · Node.js · multi-agent · CLI
Code · Docs site
IPC Vision — seven-model MNIST comparison (IPC_vision)
An exercise that walks the whole path from model to application: BP, CNN, ConvNeXt, LeNet-5, ResNet, ViT and MobileNetV2 compared under one dataset and evaluation protocol, wrapped with a Flask API, a web frontend and an Electron desktop app. The aim was to compare model scales on one task under a single evaluation protocol.
Stack: Python · PyTorch · Flask · Electron
Repository
PDF vector search (pdf-vector-search)
A local PDF knowledge-base retrieval tool for agents: documents are chunked and embedded into a local vector store, exposed through a command-line interface for an upper-layer agent to call — used for literature search and question answering. The design goal is retrieval results that carry their sources, rather than feeding papers to a model one at a time.
Stack: Python · vector search · RAG · CLI
Repository
Engineering practice
These have no dedicated repository, though they took no less time than coding work.
Paper corpus ingestion pipeline
Turning 30,000 papers into searchable, traceable knowledge entries: tiered concurrency, content-hash idempotency, and mixed local/cloud inference. It finished 30,018 of 30,031 papers in 11.2 hours (99.96%) and attached evidence links back to the source for every entry. Engineering details are in the notes.
Agent service deployment and operations
Getting a self-hosted service running on the public internet: process supervision and boot persistence, credential persistence, reverse proxy and HTTPS, private networking. Issues included a process-scoped random secret that invalidated sessions on restart and cookie scope constraints, written up in the notes.
Domestic AI chip operator exercises
Several CANN operator contest problems: 2D convolution (FP16/FP32/BF16), the HardSwish activation, and a quantised GEMM optimisation task. This involved address spaces, tiling and vectorisation at the operator level.
Private workspaces
These are research and engineering workspaces I use daily. They are private for now — email me if you would like to know more.
- Neural_Paper_Skill — a workflow skill set for neural network research and paper writing, covering ideation, training, validation, writing and figures
- CordiSwarm — a plugin-kernel research swarm agent system for long-running autonomous work
- Regain-Mamba / LesioNeXt-Code — research code for lesion analysis and remote sensing models, with reproducible training and evaluation pipelines
- WeftMesh — local AI agent software with a star topology and a device-interconnection protocol
- SwustSpirit — campus information tooling for students at my university
BearPi-Pico H3863 development archive (BearPi-Pico-H3863)
Documentation, chip manuals, firmware and source-code pointers for the BearPi-Pico H3863, archived because the material was scattered; the archive makes it easier to consult and reproduce.
Stack: embedded · documentation
Repository
HTML long-image export skill (html-png-exporter-skill)
A skill for coding agents: render a long HTML page into a previewable full-page image and export it as PNG using native Canvas. Built so that generated reports and posters can be inspected before revision.
Stack: JavaScript · Canvas · agent skill
Repository
dsh-OHDSH — an agent runtime on HarmonyOS (dsh-OHDSH)
An attempt to bring a reproducible implementation of DeepSeek Harness to HarmonyOS. Most of the effort went into diagnosing a startup failure on a real device, traced to several system-level restrictions with workarounds for each.
Stack: HarmonyOS · ArkTS · C
Repository
Other public repositories
More experimental repositories are on my GitHub profile.
