Curriculum Vitae

China

• RectifyGuard (AAAI submission, under review): formalized out-of-region content drift in e-commerce image rectification and proposed a multi-stage framework — a violation-type-aware mask refinement module (VTAM), a pixel-level paste-back mechanism, and a non-diffusion refiner (LaMa, Restormer). Introduced the EC-Compliance-50K benchmark (50K images) and outperformed Qwen-Image-Edit, Seedream 5.0, and FLUX.2 across preservation, reconstruction, and boundary-consistency metrics.
• Internal Compliance Data Management System: built a Flask-based data-governance platform with configurable scheduled pipelines that incrementally aggregate trillion-row Hive datasets into ClickHouse, plus Redis caching for millisecond-level responses, and custom computation/visualization modules for core compliance metrics.

China

• Agent Product Evaluation: co-developed a hierarchical evaluation framework for leading agent products across model capability, agent harness, and end-to-end performance, and built a weekly benchmark that surfaced capability gaps and informed product iterations.
• Business Analytics Data Agent: designed a scenario-specific Data Agent and its toolchain, integrating internal data sources, an analysis agent, and an enterprise messaging bot to enable natural-language retrieval, drill-down analysis, and report generation.

Singapore

• Intelligent Document Processing System: developed the bank-statement OCR pipeline for automated extraction of static fields and transaction tables across dozens of banks, and refactored the production data-processing pipeline to standardize heterogeneous formats.
• Credit-Risk Information Extraction: built a GPT-4.1-based workflow to extract overdue-payment information from SLIK reports, redesigning prompts and strengthening validation to reliably convert unstructured credit reports into structured risk data.

Machine Learning

• Built a complete anomaly-detection pipeline on the Freddie Mac Single-Family dataset with 68 domain-informed features structured into 7 modular feature builders.
• Compared Isolation Forest, LOF, One-Class SVM, and Elliptic Envelope; a fusion model over early-delinquency, amortization-shortfall, zero-payment, and LOF signals reached AUPRC 0.560 on the Kaggle private leaderboard vs. a 0.130 baseline.

RAG

• A LangGraph-based Corrective RAG system integrating a local PDF knowledge base with Tavily web search, using adaptive retrieval, query rewriting, document scoring, and conditional routing to reduce hallucinations.
• Processed PDFs with PyPDFLoader and recursive splitting, generated OpenAI embeddings, and built a persistent Chroma vector store with incremental loading; shipped CLI, Streamlit, and FastAPI interfaces.