/ 19 shipped projects
Shipped products, research tools, and AI experiments. Most are real projects with real users.
Most converters silently drop formulas, tables, and cross-references. any2md uses a dual-engine—PyMuPDF for text + Qwen-VL for vision—then DeepSeek LLM reconstructs context via layered prompt + few-shot + CoT. Processes 68 pages/min with 85% table structuring and 90% LaTeX formula preservation. Imported 20,000+ pages into a team knowledge base in one week, cutting manual cleanup by 70%.
Everyone assumed perceived risk would deter UAM adoption. Our TAM–SEM model on 2,609 respondents (93.1% recovery rate, KMO=0.951) proved the opposite: risk positively drives willingness to adopt (β=0.262, p<0.001). We also mined 1,695 Bilibili comments with SnowNLP to map public sentiment. The “risk paradox” finding challenged a decade of technology adoption literature. Provincial First Prize.
Enterprise-grade monorepo powering 4 apps (marketing, dashboard, docs, studio) from 6 shared packages. Next.js 16 + React 19 + Tailwind v4, with a custom design token pipeline, 541 tree-shakable icons, and multi-theme engine supporting 6 oklch color schemes. Ships features 3x faster than separate repos.
MCM/ICM 2025 (Problem B): built a system dynamics + NSGA-III multi-objective optimizer balancing tourist flow, environmental quality, and social satisfaction. Achieved 8.3% tourist prediction error with R²>0.5 on a three-dimensional coupled model. Competed against 28,000+ teams globally. Honorable Mention.
Three-model ensemble (LightGBM + Gaussian Process Regression + PSO) tested across 500+ pyrolysis data points to discover that 28.44% biomass-to-coal ratio maximizes clean energy output with R²>0.95 prediction accuracy. The entropy-weighted fuzzy evaluation framework quantified non-linear interactions between 3 yield types (gas, char, liquid). Judges praised the methodology as transferable across energy domains. First Prize + Grand Innovation Award at Shuwei Cup.
Built an integrated forecasting framework (ARIMAX-GARCH + VAR + LASSO + Prophet) spanning market size, pet population, and food manufacturing. LASSO regression identified key drivers with R²=0.9850. Cross-compared US/EU markets using HHI concentration index. Predicted 144.68M pets and 82.8% capacity utilization by 2026.
Cursor IDE extension (7 GitHub stars) that exports AI conversations to structured Markdown. Tested on 214 conversations (1,980 messages): processing time dropped from 5.8 min to 0.8 min per conversation, code snippet loss rate under 0.5%, throughput 12.5 conversations/min. Cross-platform VSIX with >98% install success rate across Windows/macOS/Linux.
Led a PEST-SWOT driven brand overhaul for Suzhou’s “林莓莓” agricultural product. Surveyed 1,200+ consumers to map purchase drivers and price sensitivity. Deployed Claude + Flux + Midjourney AIGC toolchain for brand stories and mascot concepts, cutting creative iteration from weeks to days. Projected 40%+ user engagement increase. National Second Prize.
Open-source Claude Code skill that wraps MinerU for layout-aware PDF/document conversion. One-command install, no API keys, 3 supported formats (PDF, DOCX, PPTX). Handles 100+ page documents with table and formula preservation. Published on Smithery marketplace, 5+ community installs in first week. Built for the Claude Code ecosystem with auto-detection of document type and intelligent chunking.
Multi-rotor drone system for automated warehouse stocktaking. Integrated STM32 flight controller + OpenMV vision + BLE communication for autonomous waypoint navigation and QR code scanning across 3D shelf structures. 95% barcode recognition accuracy with laser calibration, completing full inventory traversal in under 3 minutes.
Turned ad-hoc prompt engineering into a structured agent protocol: task decomposition, state machines, retry strategies, and auditable intermediate states. Tested on 120 cross-domain tasks: success rate jumped from 63% to 81%, JSON compliance from 72% to 94%, rework rounds dropped 29%. Framework-agnostic—plugs into LangGraph, Semantic Kernel, or custom executors.
Every codebase accumulates Vibe Coding debt: custom date formatters, DIY loggers, bespoke state machines. Next-Unicorn audits your code via Context7 MCP, identifies reinvented wheels, generates migration plans, and ships delete-code checklists. 176 tests passed with 29 property-based verifications. Published on Smithery + npm, supports 35+ AI agents including Claude Code, Cursor, and OpenCode.
React + TypeScript monitoring system tracking 12 water quality parameters (pH, turbidity, dissolved oxygen, etc.) with 5-second auto-refresh and configurable alert thresholds. Built with streaming data visualization and historical trend analysis for environmental monitoring.
Production-grade automated annotation system for document page elements. Async 4-stage pipeline with Qwen VL models identifies 50 element types (12 code languages, 13 interaction formats, 12 content elements, 13 other tags). Three-layer fault tolerance (Retry + RateLimit + CircuitBreaker), SHA256 content-addressed cache, and real-time SSE monitoring dashboard.
Multi-modal visual search relevance labeling pipeline using LLMs. Processes image + text signals across 8 annotation dimensions (visual relevance, content relevance, query/doc drop, functional parity, category granularity, visual similarity, modal corroboration). 19 prompt versions with automated comparison and reflection analysis tools.
Multi-platform video content analysis system covering YouTube, Instagram, TikTok, VK, and Facebook. Refactored from legacy monolith to modular architecture: 5 platform fetchers, 6 batch processors, 5 video analyzers. Fast text analysis achieves 90%+ accuracy with 460x acceleration over manual review, processing 1,207+ videos across batch runs.
Open-source quantitative trading copilot for crypto markets. 8 microservices + 9 infrastructure services (17 containers total) with unified TUI cockpit, 41 technical indicators, real-time market data streaming, and in-context AI copilot. TypeScript 5.7 + Python 3.11+, CI/CD with GitHub Actions.
Comprehensive quality evaluation framework for the Habitat-Matterport 3D Research Dataset (1,000 real 3D scenes). Assessed 15+ dimensions including mesh quality, texture integrity, semantic annotation accuracy, and metadata completeness. Achieved 4.62/5 overall score with 100% acceptance pass rate (8/8 criteria). 95% requirement compliance.
Built multi-tenant reporting infrastructure for Appen (NASDAQ: APT). Next.js 16 monorepo (5 apps + 10 packages) serves TCS, ByteWorks, and DataPower dashboards with Upstash QStash async job queues and Pusher real-time. The breakthrough: a production-grade demo mode with MSW + tRPC mock interception, privacy redaction pipeline (schema validate → redact → normalize → value protocol), and 4 behavioral strategies—solving the #1 enterprise SaaS sales problem.
Meta-labeling pipeline that audits and corrects mislabeled AI-generated video training data using Gemini 2.5 Flash. Classifies 6 attributes (quality, gender, age, race, scene, season) per video with EWMA-adaptive rate limiting and HTTP/2 concurrency. Processes 60MB+ videos via Gemini Files API without local download. 3 iterative rounds with incremental merge. Human labeling: ~$1/video. This pipeline: ~$0.01/video at scale.
Production URL consistency pipeline that defeats TLS-fingerprint bot detection at zero marginal LLM cost. curl_cffi spoofs Chrome 124's full ClientHello (cipher suites, extensions, ordering)—not just User-Agent. Playwright renders JS-heavy SPAs. Local Qwen3:1.7b via Ollama compares semantics without API fees or GDPR exposure. Rotating-ad detection re-fetches 3× to flag domain changes. 30 HTTP + 5 Playwright + 5 LLM concurrent.
First systematic vendor evaluation framework for 3D Gaussian Splatting content production. Multi-round scored assessment across 6 dimensions: HP/LP geometry (pivot + retopology), PBR texture completeness (BaseColor/Normal/Metallic/AO), file naming conventions, semantic annotation, preview standards (8+ renders), and communication SLA. The revelation: vendors weren't technically incapable—they were operationally immature. Multi-round CSV tracker makes every revision cycle auditable.
