handa wang

Crypto Market Intelligence Platform — Time-Series Forecasting, Feature Engineering & Champion-Challenger Evaluation
• Built an end-to-end crypto market data science pipeline for Bitcoin market prediction, covering OHLCV ingestion, orderbook snapshots, funding rates, volatility indicators, feature generation, model training, offline evaluation, and experiment tracking.
• Designed leakage-safe supervised learning datasets from multi-horizon time-series windows, including lag features, rolling statistics, technical indicators, volatility-regime features, return normalization, missing-data checks, and timebased train/validation/test splits.
• Trained and compared statistical baselines, LightGBM/XGBoost models, and PyTorch sequence models for multihorizon price-direction and return-regime prediction using walk-forward validation and time-series cross-validation.
• Built a champion-challenger evaluation framework to compare baseline and experimental models under identical forward-market windows, transaction-cost assumptions, and risk constraints before simulated deployment.
• Evaluated models with ML and market-oriented KPIs, including AUC, F1, precision@K, hit rate, information coefficient, Sharpe ratio, max drawdown, and transaction-cost-adjusted simulated returns, and built a Power BI reporting layer connected to PostgreSQL/exported outputs for model-performance, risk, and champion-challenger dashboards.
• Implemented feature-importance and error-analysis workflows using SHAP, permutation importance, regime-based slice analysis, and prediction drift monitoring to analyze model behavior under high-volatility and low-liquidity market conditions.
• Stack: Python / pandas / NumPy / scikit-learn / LightGBM / XGBoost / PyTorch / PostgreSQL / Power BI / Docker / Linux

Insurance Industry — RAG-based Failure Investigation Assistant & Test Automation Platform Sub-leader in an 8-member team / 40+ total project members
• Contributed to an enterprise test automation and failure-analysis platform, integrating deterministic E2E test workflows with LLM/RAG-assisted investigation for core insurance-system failures.
• Led common-module refactoring by separating low-level page operations from business-level actions, increasing common-module reuse to 80% and reducing test-development lead time by 75%.
• Designed a RAG-based failure investigation assistant that retrieved historical Redmine tickets, known failure patterns, test documents, troubleshooting notes, and sanitized logs to support root-cause hypothesis generation.
• Implemented a safe LLM-assisted workflow with sanitized failure context, retrieval-grounded evidence, structured output review, and human confirmation before ticket updates or final root-cause judgment.
• Stack: Python / Java / Selenide / PostgreSQL / ChromaDB / OpenAI API / Pydantic v2 / Redmine API / AWS / Windows