Projects

I turn research questions and practical needs into systems that can run, be verified, and remain maintainable.

This page records my research engineering, local intelligent tools, and software systems, with emphasis on the problem, my contribution, the current result, and the public entry points.

View representative projects GitHub profile

This page includes only work with public code, documentation, or reviewable outcomes. Unpublished material, private data, and unapproved experimental results remain outside its scope.

Representative cases

Three representative project cases.

These cases cover research control, research knowledge management, and a public delivery system, each organized around problem, contribution, and result.

Phase completed Mar 2025—Jul 2026 Independent developer

Research engineering

CBF-Rover: Safe Control for Mobile Robots under Stochastic Disturbances

For stochastic disturbances and dynamic obstacles, this project combines reinforcement learning, sparse polynomial controllers, and stochastic barrier-certificate optimization with online CBF action correction and batch closed-loop simulation.

Problem: Learned policies must balance target reaching, obstacle avoidance, and verifiability. The robot also needs timely corrections to actions that could violate safety constraints under disturbances and moving obstacles.
My work: I independently developed policy training and optimization, dynamics models, and sparse polynomial approximations. Stochastic barrier certificates (SBCs) and SOS/SDP optimization compute probability lower bounds and provide feedback for controller improvement. Python and MOSEK Fusion support batch disturbance and obstacle-trajectory simulations, alongside online CBF safety filtering.
Phase results: Simulations achieved safe target reaching and obstacle avoidance across multiple scenarios, while research probability lower bounds approached Monte Carlo estimates through iteration. Related methods are published in my first-author MEMOCODE 2026 (a constituent conference of ESWEEK 2026) paper, with follow-up work submitted to a CCF Class A conference. Robot-side CBF behavior is checked separately in numerical dynamics simulation and Gazebo; these empirical results are not a formal probabilistic guarantee for the complete robot system.

PythonMOSEK FusionSOS / SDPOnline CBF

View repository Architecture Public records

01
Reference policy and modelReinforcement learning · dynamics
02
Polynomial controllerSparse approximation for verification
03
Probability lower boundSBC · SOS / SDP optimization
04
Certificate feedbackController improvement · Monte Carlo comparison

Models and certificates support research probability lower bounds. Online CBF correction and Gazebo simulation provide separate evidence of robot closed-loop behavior.

In active iteration Jul 2026—present Independent design and development

Research infrastructure

Research Memory: Local Research Knowledge Retrieval

This local knowledge layer organizes structured sources, paper notes, code records, and experiment provenance, then exposes read-only retrieval to Codex MCP and uTools.

Problem: Sources, code changes, and experiment outputs live across directories and tools, while ordinary full-text search rarely preserves provenance, version, and experiment relationships together.
My work: I designed records with content hashes and provenance fields, separated structured data, FTS5, vector indexes, and runtime cache, and exposed four bounded read-only paths for search, stable-ID opening, metric queries, and service status.
Current result: The system now provides a runnable local retrieval service, Codex MCP tools, and a uTools entry point. The public repository includes code, architecture, templates, and tests, but no personal sources or database.

PythonSQLite FTS5OpenVINOMCP

View repository System architecture

01
Authorized sourcesDocuments · notes · code · experiment records
02
Structured ingestContent hashes and provenance
03
Hybrid indexesSQLite FTS5 · vector index
04
Read-only gatewayCodex MCP · uTools

The data layer stays separate from public code, and external tools cannot write directly into the knowledge store.

Live 2024—present Independent design and maintenance

Public academic system

Bilingual Academic Portfolio

Built with vanilla HTML, CSS, and JavaScript, this bilingual academic portfolio uses structured data, a zero-dependency validator, and a separate analytics service to keep releases maintainable.

Problem: Research, projects, CV evidence, and public materials need equivalent purpose and reliable navigation across Chinese and English, while remaining searchable, mobile-friendly, and maintainable.
My work: I built the bilingual pages, responsive navigation, dual themes, lightbox, SEO, and JSON-LD contracts, then wrote a Node.js validator for resources, structured data, fact inventories, and critical breakpoints.
Current result: GitHub Pages publishes the site directly, while Cloudflare Worker + D1 hosts analytics separately. The pages, validator, and delivery process are publicly reviewable.

HTML / CSS / JavaScriptGitHub PagesNode.js ValidationCloudflare Workers / D1

Open site View source Public architecture

01
Chinese and English pagesEquivalent purpose and public entry points
02
Shared assets and behaviorCSS · JavaScript · themes · keyboard
03
Semantic validationSEO · JSON-LD · Node.js validator
04
Separate delivery surfacesGitHub Pages · Worker / D1

Static pages and analytics ship independently; committed files are the public GitHub Pages input.

Local tools

Local AI and knowledge tools.

These tools grew from my own reading, media, and content-organization workflows, with their current state and human decision boundaries kept visible.

MaintainedBrowser and document clients

LocalReadTranslate: Local Selection Read & Translate

Across browsers, Word, and WPS, the tool captures only text the user selects, sends it through a local service for reading, translation, and formula handling, and returns reusable results.

Path: Selected text → browser script or document client → local FastAPI → Kokoro / Ollama → speech, translation, or LaTeX.
Boundary: Processing is local by default; selected text reaches a self-configured remote model only when that mode is explicitly enabled.

FastAPIBrowser ClientKokoro TTSOllama

View repository

Usable workflowMedia processing

Bilingual Subtitle Pipeline

The pipeline selects from sidecar or embedded subtitles, PGS image subtitles, and Whisper transcription, then produces Chinese–English ASS after correction, translation, and timeline validation.

Path: Source analysis → extraction / OCR / ASR → source correction → translation → timeline restoration and validation.
Result: A local web console, checkpoints, resume support, and source-language, Chinese, and bilingual ASS outputs are implemented.

PythonFFmpeg / WhisperPGS OCROllama

View repository

In developmentLocal imaging workflow

Local AI Photo Selection Workbench

The workbench combines quality pre-screening, similar-photo clustering, and pairwise comparison to provide local selection guidance while reserving final decisions for the user.

Path: Quality pre-screening → similar-photo groups → pairwise arena → human confirmation, with JPG / HEIC and RAW counterparts kept together.
Boundary: AI provides observations and ordering suggestions; disputed photographs remain for human judgment rather than an automatic aesthetic verdict.

FlaskOllama / QwenDINOv2RAW / HEIC

View repository Architecture

Customized and maintainedUpstream fork

BiliClaw Extended: Local-First Content Discovery

Built as a customized OpenBiliClaw fork, this local-first multi-platform discovery flow adds account isolation, recoverable indexing, tab reuse, and explicit profile-update control.

My changes: Current work focuses on account-isolated Douyin bootstrap, recoverable liked-item vector indexes, batch tab reuse and reinjection, and separation of task confirmation from profile updates.
Attribution: This is a customization of OpenBiliClaw, not a claim of authorship over the upstream system.

FastAPIBrowser ExtensionSQLite / ChromaDBOllama

View customized fork Multi-source design

Software engineering

Software engineering practice.

Two complete undergraduate systems, retaining their requirements, my role, deployment shape, and reviewable outcomes.

CompletedUndergraduate capstone · Dec 2022—Jun 2023Project lead

Multi-Platform Attendance System Using Face Recognition

The system identifies students from camera-captured or uploaded images and produces attendance records, summaries, and periodic reports across Android, WeChat Mini Program, and Web clients.

Problem: Reduce the time cost and proxy-attendance risk of roll calls by connecting recognition, attendance records, and reports across multiple clients.
My work: I independently implemented the client interfaces, business backend, image-processing path, and end-to-end data workflow.
Result: The capstone received Second Prize for Outstanding Undergraduate Thesis (Design), Southwest University Class of 2023; its code and full thesis are public.

uni-appSpring BootMySQLPython

View repository Read thesis View award record

CompletedChinasoft International training · Team project · Jun—Jul 2022Project Manager & Technical Manager

MicFamily: KTV Operations Management System

For day-to-day KTV operations, the system centralizes staff, member, room, order, product, and notice management; customers can look up orders, reserve rooms, settle online, and purchase products.

My work: I coordinated task breakdown, technical decisions, schedule, front-end/back-end integration, and project delivery.
Implementation: The browser UI uses HTML/CSS, Vue, jQuery, Bootstrap, and Axios; the service layer uses Spring Boot, MyBatis-Plus, and MySQL. Static front-end assets and HTTP APIs ship in one Spring Boot application.
Result: Delivered the training project, was recognized as an Outstanding Intern, and received a university internship score of 94.

VueSpring BootMyBatis-PlusMySQL

View repository Project report Internship record

Development infrastructure

Developer infrastructure.

Turning provenance, versions, and recovery steps that tend to drift across personal development environments into a verifiable public tool.

MaintainedDevelopment environment

Reproducible Codex Skills Kit

Tracks Codex Skill and plugin provenance, pinned versions, and states, with cross-platform installation, validation, and recovery workflows.

Focus: Pin source and version inventories, then use installers, status checks, CI, and recovery steps to reduce configuration drift across machines.
Result: The public repository supports review of configuration provenance, installation reconstruction, and verification of the restored component inventory.

PythonPowerShellCodex SkillsCI

View repository