Technical Approach: Developed a Windows tool for USB connection management and firmware upgrades for VR headsets (Android/Linux based). Leveraged WinUSB/SetupAPI for device enumeration, driver loading, and USB communication; built ADB modules for mode switching, status query, and log capture; designed Fastboot upgrade with partition parsing, chunked writing, CRC validation, and auto-retry with breakpoint resume. Collaborated with firmware team to optimize adbd startup timing and Linux USB Gadget configuration.
Challenges: ① Driver binding conflicts – resolved by modifying INF files to match VID/PID and using SetupAPI to dynamically load drivers. ② USB mode switching failure – analyzed sysfs and Android property mechanisms to fix adbd startup sequence, proactively sending mode switch commands. ③ Upgrade interruptions – added chunked transfer, CRC checks, and breakpoint resume; auto-retry on timeout.
Results: Device recognition rate improved from 85% to 99%, recovery time from 2 minutes to <5 seconds; firmware upgrade success rate from 90% to 98%, supporting 500+ units/day on production lines; delivered a troubleshooting manual covering 15 common failure scenarios, improving after-sales efficiency by 40%.

Technical Approach: Built a Playwright-based automation framework to accurately capture declaration data (header, body, receipt number, PDFs) from the Single Window platform. Supported multi-terminal synchronization and fast historical data retrieval.
Challenges: ① Dynamic DDT (Data Token) link extraction – executed custom JavaScript via page.evaluate() to obtain real-time DDT values. ② Encrypted table body decryption – called decryption functions to decode data. ③ Dynamic table creation – used ALTER TABLE to auto-adjust schema based on varying API response fields.
Results: Data accuracy ≥99.5%, automation efficiency improved by 70%, historical query time reduced to minutes, dramatically reducing manual workload.

Technical Approach: Developed an OpenCV-based automatic slider CAPTCHA solver. Offline: used adaptive Gaussian threshold, contour extraction, area filtering (0.5~1.5× gap area), and Hu-moment matching to locate the gap. Online: designed a real-time feedback loop using mss screenshots, dynamically adjusting drag displacement (1/1.2 of remaining distance + random vertical offset) until gap <3px.
Challenges: ① Gap localization accuracy – filtered contours by area and ranked top 3 by Hu-moment similarity to reduce false matches. ② Real-time performance – each cycle completed in ~0.3s, convergence within 9 iterations. ③ Human-like trajectory – used non-linear steps and random jitter to avoid detection.
Results: Success rate ≥90%, average time ≤3 seconds, reduced manual CAPTCHA cost by ~60%. The approach is portable to various slider styles.

Technical Approach: Led the architecture and development of an HS classification agent system based on "Multi-modal input → Rule MCP generating candidate codes → Dual-source RAG retrieval → LLM decision". Designed a multi-agent collaboration framework including Rule Agent, Case/Regulation Retrieval Agent, Reasoning Agent, Counter-evidence Agent, and Review Agent, orchestrated by a Scheduling Agent. Built a weighted dual-source RAG (authoritative case base + tariff regulation base) with dual-vector recall, low-similarity filtering, and positive/negative clause retrieval. Implemented hard validation of GIRs, four-dimensional scoring (Rule 40% + Cases 30% + Clauses 20% + Principles 10%), counter-evidence reasoning, and a human review feedback loop for incremental knowledge base updates.
Challenges: ① Candidate code omission – designed three-tier rule execution (mandatory + industry-specific Python rules + fallback), with MCP timeout circuit-breaking and 3× exponential backoff retry; candidate omission rate ≤3%. ② RAG noise – used dual-vector recall with feature weighting, filtered low-similarity (<0.6) cases, and applied reranking to reduce noise by ≥90%. ③ Complex goods reasoning – employed multi-agent collaborative reasoning with counter-evidence and review mechanisms; auto-triggered secondary verification for low-confidence results, achieving ≥80% accuracy for complex goods.
Results: Accuracy ≥92% for standard goods and ≥80% for complex goods; human correction rate ≤8%; high-risk documents ≤15%. Supports real-time classification (P95≤2s) and batch processing (≥50 items/min) with high concurrency, fault tolerance, and full logging, significantly reducing manual workload while meeting customs compliance.