Building reproducible tools for audit analytics, credit risk, and quantitative research.
关注审计数据分析、信贷风险与量化研究，把方法做成可以运行、检查和复现的项目。
An audit analytics workflow for a synthetic 30,000-voucher ledger: nine review procedures, Benford analysis, Isolation Forest, an explainable risk score, a queryable SQLite warehouse, a six-page reviewer dashboard, and a budgeted review workpaper with validated human outcomes.
Start here: Source & setup · Methodology and limitations · Review workpaper · Reproducible reports
Its injected anomalies match the rule definitions; the benchmark measures the workflow, not real-world fraud detection.
A point-in-time credit cohort workflow: application-time features, fully mature 180-day delinquency labels, separate training and calibration vintages, untouched out-of-time testing, early score monitoring before labels mature, and reviewable static reports with downloadable rows.
Start here: Live synthetic report · Early score monitor · Methodology & run instructions · Predictions · CI runs
All published data is synthetic. The independent calibration worsened Brier in this run; the negative result is reported rather than tuned away.
A Python research pipeline connecting factor evaluation, walk-forward modeling, portfolio construction, transaction costs, and HTML reports. Includes a FastAPI service and a Streamlit dashboard.
Start here: Documentation · Computed sample · Source & tests
The bundled sample data is synthetic; its results demonstrate the workflow and do not establish a tradable edge.
An event-driven market microstructure and execution research platform with a C++20 book core and Python reference model. Its published cost comparisons use synthetic sessions and report both arrival-price and interval-VWAP benchmarks.
Explore: Reproducible demo · Design and test evidence · Transaction cost guide
A double-entry ledger for research fills with exact USD accounting, deterministic replay, a hash-chained JSONL journal, and optional external head checkpoints. It shows cash, historical inventory cost, and realized P&L; the synthetic demo does not claim market valuation or live trading performance.
Explore: Accounting model · Reproducible demo · CI runs
| Project | Focus |
|---|---|
| FactorLab · A-share Multi-Factor Research | Live report, 中文文档, and purged evaluation. The default run is synthetic; real-market point-in-time statement normalization is still in progress. |
| High-dimensional Causal Allocation Lab | Browser demo for simulated covariate adjustment, causal estimation, and robust allocation. |
| Interval Financial Risk | Experiment report comparing point and distributional features, with temporal validation and downloadable predictions. |
| Investor Network GNN | Three-seed benchmark: fixed-checkpoint graph ablations, saved predictions and independent metric checks. |
| WeCom Agent Platform | Document retrieval and query workflows with a Python backend and React interface. |
Personal AI Chat also provides a local simulated chat mode and a separately configured private model connection.
| Project | Stack |
|---|---|
| AuditLens | Python, pandas, scikit-learn, SQLite / SQLAlchemy, Streamlit |
| CreditVintage | Python, NumPy, scikit-learn, HTML/CSS, pytest, Ruff, mypy |
| AlphaForge | Python, LightGBM, CVXPY, Parquet / DuckDB, FastAPI, Streamlit |
| TradeForge | C++20, pybind11, Python, Parquet / DuckDB, FastAPI, Streamlit |
| LedgerX | Python standard library, Decimal, pytest, Ruff, mypy |
Across the other projects I also use PyTorch, TypeScript and React.
I focus on explicit data provenance, reproducible experiments, baseline comparisons, and tests that check model and application behavior. Each repository documents its setup and current limitations.