Tim Zhang

Problem

Hong Kong Architectural Services Department (ArchSD) case processing relied on manual entry and routing, with average processing cycles lasting days and high error rates.

Approach

• Designed 7-layer Agentic processing architecture with 17 pluggable atomic capabilities
• Built multi-format document parsing pipeline with field extraction (A-Q categories) for ICC 1823, TMO, RCC channels
• Integrated OpenAI GPT-4o with text-embedding-3-small for intelligent classification and response generation
• Implemented pgvector-based semantic case retrieval with RRF fusion ranking
• Developed 3-tier quality assessment system (L1 keyword → L2 rules → L3 RAGAS LLM-as-Judge)
• Created self-healing mechanism with Best-of-N selection and 4-category failure differential rollback
• Deployed on Google Cloud Run with Cloud SQL and Docker for auto-scaling

Outcome

67 commits, 80,430 lines of code across two iterations. 12/12 requirements fully implemented with end-to-end automation. Reduced processing time by 75% (40 min → 10 min). Projected annual cost savings of HK$6.39M.

Problem

University students, faculty, and administrators struggled with dispersed policy documents across 30+ departments, leading to 70%+ retrieval failure rates and inconsistent information delivery.

Data

Ingested 200+ policy documents, academic regulations, administrative guidelines, and FAQ corpora across three user personas (student, faculty, admin).

Approach

Built AgenticRAG with Hybrid Retrieval (dense + sparse + reranking), Adaptive User Memory for personalized dialogue, RBAC-based access control, and LLM-as-Judge for answer quality assurance.

Outcome

Policy Q&A accuracy >95%. Average retrieval time reduced by >70%. System response within 10 seconds. Demonstrated at Lingnan Innovation Showcase.

Problem

Walmart China's B2B financial reporting relied on fragmented data sources across SQL Server, StarRocks, and legacy systems, causing delayed reporting, undetected data anomalies, and manual reconciliation overhead.

Data

414,000+ contract records, monthly B2B transaction partitions (9,000–10,000 records/month), cross-source financial metrics including order amounts, fulfillment types, and credit ledger balances.

Approach

Built cross-source financial data quality monitoring with automated anomaly detection (e.g., fulfillment amount exceeding order amount). Designed B2B reporting iterations covering contract lifecycle, countdown tracking, and credit ledger reconciliation. Implemented Feishu alerting for downstream BI consumers.

Outcome

Reduced data incident detection latency from days to minutes. Achieved automated cross-source validation across StarRocks and SQL Server. Enabled consistent financial reporting for B2B operations across multiple business units.