Knowledge Base

Multimodal AI virtual try-on combining diffusion generation, YOLO12n-LC classification, and LLM agent control via Dify workflow orchestration.

Metrics: fid: 28.5 (19% improvement), classificationAccuracy: 94.2%, modelSize: 8x smaller than YOLO12n, userSuccess: 87%

Industrial IoT smart manufacturing platform deployed across 10+ factory sites. Spring Cloud microservices + Android shop-floor apps + Vue.js portals + IoT (RFID, scales, conveyors).

Metrics: factories: 10+ sites, efficiency: 30%+ improvement, workers: Hundreds daily users, team: 6-person cross-functional team

Enterprise IM platform (Android/Web/PC) serving 5,000 DAU over 10 years. Started as sole developer, grew to 20-person team. Custom TCP protocol (WebSocket unstable), NDK TCP/UDP for sub-200ms latency. Evolved from standalone IM to multi-subsystem platform. Migrated C++ messaging to Easemob cloud, eliminating 90%+ defects. Still in use and actively growing.

Metrics: dau: 5,000 daily active users, latency: <200ms (custom TCP protocol), messages: 500K+ per day, downtime: <2% over 10 years, evolution: Sole developer → 20-person team, status: Still active and growing

7-year IoT product line: smart switches and gateways sold to customers. Sole Android developer for all mobile apps. Gateway firmware for Zigbee/WiFi → MQTT. Participated in architecture design (Java Spring, time-series DB, message queue). 20-person team with majority backend.

Metrics: duration: 7 years (2016-2023), teamSize: 20 (majority backend), products: Smart switches and gateways sold to customers, layers: Mobile + Gateway + Backend

Evaluated and participated in an Android hotfix framework using method replacement (AndFix) and Dex merge (Tinker). Identified critical limitations: method-only replacement, no native .so swapping, ROM compatibility gaps, unsustainable version management. Led the evaluation that resulted in retiring the system in favor of staged rollout + feature flags.

Metrics: evolution: AndFix → Tinker → retire → staged rollout + feature flags, decision: Retired after evaluating maintenance cost exceeded business value

Full-stack streaming platform: Android client (C++ core via NDK), ASP.NET web admin, WebSocket chat. RTMP/HLS delivery, multi-language architecture spanning 6+ languages.

Metrics: 3-year production run, multi-language architecture (6+ languages)

Offline-first Android GIS app for forest rangers in no-signal areas. Custom offline maps, GPS tracking, incident reporting with deferred sync. Supports risk event reporting for fire hazards, pest/disease, and land occupation.

Metrics: Deployed for forestry department, reliable zero-connectivity operation

School face recognition attendance with local on-device comparison using ArcSoft SDK. 3-person team — I built the Android terminal with face recognition, liveness detection, and gate control. 1-2s per student, 99%+ accuracy, real-time parent notifications. Deployed across multiple schools.

Metrics: recognition: 1-2 seconds per student (local comparison), accuracy: 99%+, schools: 3+ deployed, notification: Real-time parent push

Smart power platform for factories and buildings. 20-person company-wide project. Monitor per-load electrical parameters, remote control, alarm policies. 15% energy saving, sub-second alarm, 99.9% data collection rate, 3+ parks deployed.

Metrics: energySaving: 15% average reduction, alarmSpeed: <1 second, dataAccuracy: 99.9% collection rate, parks: 3+ deployed, teamSize: 20 people

Multi-terminal platform: admin booking, face recognition terminals, visitor WeChat mini-program, WebRTC sessions.

Metrics: 99.5% face recognition accuracy, <1s verification, 30% efficiency gain

Smart library locker with Android control cabinet (locks, UV, sensors) + borrower app + face/QR authentication.

Metrics: <3s borrowing, multiple schools, 0% book loss rate

Smart gateway controlling non-networked circuit breakers via RS232. Each gateway manages up to 70 breakers. Telemetry, remote control, local alerting, cloud reporting. Connects to Smart Power Management Platform. 4-person team — I developed the Android gateway and designed the architecture.

Metrics: capacity: Up to 70 breakers per gateway, telemetry: Real-time current/voltage/power/energy, availability: 24/7 edge operation, integration: Connects to Smart Power Platform

Manufacturing information systems for a high-security banknote paper production facility. Designed gateway to aggregate data from 7-8 legacy hardware types via serial ports (9600/115200 baud). Synchronized scattered databases to unified BI layer (FineBI). Deployed UHF scanners for paper axis tracking despite metal wire signal interference.

Metrics: hardwareDevices: 7-8 types, ~12 devices, 2 gateways, serialBaud: 9600 / 115200, reportingInterval: ~30 seconds, biTool: FineBI (帆软), siteApproval: ~1 week lead time, participation: Intermittent

Intelligent service robot for exhibitions and malls. Robot provided SDK for control; team built features on top: navigation, product introduction, face recognition with check-in, knowledge base, and voice interaction. 4-person team — I developed the Android interaction interface and integrated robot SDK features. Deployed at client exhibition, still in use.

Metrics: navigation: Centimeter-level precision (LiDAR SLAM), faceRecognition: >98% accuracy, voiceUnderstanding: >90% intent understanding, availability: 7x24 continuous operation, deployment: Client exhibition, still in use

Digital Signage platform for centralized ad delivery across 1000+ retail display devices. Three-component architecture: Vue.js Admin Console, Customer Portal, Spring Boot/Cloud backend, Android Device Client. MQTT broadcast with high-concurrency handling, multi-screen adaptation, campaign scheduling, device fleet management, and offline cache/recovery.

Metrics: devices: 1000+ display devices, teamSize: 9 (3 backend, 2 frontend, 2 mobile, 1 designer, 1 QA), deliveryTime: 2 months, ahead of schedule, communication: MQTT broadcast with high-concurrency handling, businessScenario: Retail stores (bookstores, clothing stores, design shops)

Cryptocurrency trading platform mobile app with real-time market data via WebSocket, coin search, exchange operations, and secure wallet management.

Metrics: Real-time market data updates, secure wallet with encryption

General-purpose AI platform for data annotation and model training. Validated with Chinese herbal recognition demo (20 categories, ~50 images/class). Prototype → handoff → launch, but not maintained afterwards.

Metrics: demoCategories: 20 herbal types, imagesPerClass: ~50, modelsTested: 4 (YOLOv4, ResNet, MobileNetV3, EfficientNet), evaluationCriteria: Top-1 Accuracy, Macro F1, Worst-class Recall, P95 Latency, outcome: Developed and launched, not maintained

AI platform for equipment predictive maintenance. Uses historical logs, fault records, and runtime indicators to model aging patterns and predict maintenance schedules. Outputs risk scoring and data-driven scheduling.

Metrics: Multi-model comparison with aging trend analysis and risk scoring

Vehicle inspection system for highway toll stations. C# application integrating electronic scales for vehicle weighing, real-time data storage, overload alarm alerts, and customer system integration for automatic invoice generation.

Metrics: Deployed at highway toll station with weighing, alarms, and invoicing

Early-career solo project: built entire enterprise patent management system myself — .NET desktop app with Lucene.NET full-text search, SQL Server, Excel batch processing. 10,000+ patents, <2s search, 50+ daily users.

Metrics: patents: 10,000+ managed, search: <2s response time, users: 50+ daily users, team: Solo project