skills

Listed by actual level of use, without inflation. Every line corresponds to something I have built.

DEEP LEARNING AND DATA

Area	Level	What I have done
PyTorch training and evaluation	Comfortable	Classification and segmentation training, external validation and metrics; fixed splits, seeds and evaluation scripts
Medical imaging data	Comfortable	Ultrasound and CT preprocessing, augmentation and ROI handling; unified evaluation across datasets
Interpretability	Comfortable	Grad-CAM attention visualisation to support readable model decisions
Model comparison and selection	Comfortable	Seven classification models compared under one dataset and protocol; segmentation model selection for medical imaging
Remote sensing experiments	Some experience	RGB-T adaptive image fusion and budgeted semantic segmentation: experiment design and ablations
Lightweight and on-device deployment	Some experience	Model export, inference packaging and CPU-fallback paths
NLP / small language models	Beginner	Small-model inference and fine-tuning experiments

Multi-agent orchestration: role-based task splitting (explore, implement, review) with parallel execution; progress on disk and resumable long-running work
Own coding agent: incremental changes on an open-source base adding persistent memory, multi-agent workflows for smaller models and goal-driven execution; published to npm with its own documentation site
Client protocol design: an agent-bridge protocol for mobile (21 requests, 18 events) covering streaming output, long-task progress and session recovery
MCP services: a paper knowledge base exposed as vector retrieval plus an MCP service on an edge platform, usable from any client
Plugins and runtime changes: reasoning-mode routing via injectors, cross-device memory sync, session main-loop switching with the built-in loop as fallback, and packaging an existing workflow suite as a plugin
Skill module design: 13 modules plus about 20 deterministic scripts and 90+ tests for a research workflow suite; 9 modules for a hardware design pipeline
Device mesh: star topology with a purpose-built device protocol supporting capability advertisement, task dispatch and reconnection (three platforms, 176 passing tests)
Process governance: project identity, goals and task execution constrained by a plain-Markdown specification

Linux services: system-level and user-level systemd supervision, boot persistence, log triage
Service exposure: reverse proxies, private networking, tunnels and custom domains, HTTPS and security headers
Cloudflare platform: Workers, vector and key-value storage, Pages, tunnel and DNS record management
Continuous integration: dual-target publishing (static hosting plus edge platform) from GitHub Actions, build caching, headless-browser artifact generation
Model serving: an out-of-tree vLLM plugin (model implementation, configuration and tool-call parser); local llama.cpp CUDA inference services
Cross-device sync: memory and experience synchronised through a private Git repository
Frontend and clients: Vue3 + Three.js + Electron desktop application; Flutter Android application
Testing and delivery: full-interface automated regression (13 pages, 265 buttons, 192 actual clicks), release packaging and offline delivery, verification on other machines
Retrieval and knowledge bases: vector search, document chunking, evidence trails and review queues
Fonts and frontend engineering: font subsetting, device-matched font stacks, post-build artifact cleanup
Python engineering: uv / venv environments, scripts and CLI tools
Git: branching and rebasing, history cleanup, repository mirroring

PCB design: schematic and PCB layout of an STM32F407 minimum system board in LCEDA, with a ready-to-order BOM
Flashing and debugging: Keil toolchain, flashing port and build-cache troubleshooting
Embedded platforms: working with development material for BearPi-Pico H3863 (WS63) and similar boards
Domestic AI chip operators: CANN operator implementation and performance work (convolution, activation, quantised GEMM)
Digital design: Verilog and RISC-V processor design study (RTL simulation stage)

Record-keeping: criteria set before the run; failed runs recorded in the same template as successful ones; configurations and conclusions kept in a registry
Reproducibility: splits, seeds and evaluation scripts in the repository so results reproduce elsewhere
Evaluation metrics: the standard medical imaging set (AUC, sensitivity, specificity, precision, F1) and what each does and does not mean clinically
External validation: independent test sets preferred over comparisons on the same source data

Maintains a bilingual site; full technical documentation experience (architecture notes, interface documentation, deployment steps)
Multi-language READMEs for open-source projects (including one covering ten languages)
Project review material: technical proposals, experiment records, presentation decks and scripts

Toolchain: Linux for development, uv for Python environments, Git and GitHub for code, self-hosted servers for deployment
Working with agents: role-based splitting with parallel execution; a rules file and memory file per project; recurring problems upgraded into process constraints — see the notes
Experiments: criteria fixed before the run; failures recorded on equal footing with successes
Delivery: everything machine-checkable goes to scripts; anything requiring human judgement is confirmed by a person before delivery