Zetfung Liu 劉哲鴻
Welcome to my homepage. I study economics at the Central University of Finance and Economics, and I am interested in quantitative finance and data-driven research. I keep my notes in Chinese on the blog.
About Me
| Hometown: | Jingdezhen |
|---|---|
| Human Languages: | Mandarin (native), Hakka (native), English (CET-6 604, no prep), Cantonese (basic) |
| Computer Languages: | Python |
| Interests: | Calligraphy, Reading, Table Tennis |
| Email: | zhehongl91[at]gmail[dot]com |
Education
| 2024–2025 | Central University of Finance and Economics, Public Finance and Taxation | Undergraduate |
| 2025- | Central University of Finance and Economics, Economics | Undergraduate |
Internships
| 2026.11- | Western Securities, industry research | Research Intern |
Projects
ccb: Multi-Provider Terminal Coding Agent2026.04 – 2026.08
- Ownership: Independent derivative of Claude Code (not affiliated with Anthropic), rebuilt from the v2.1.88 sourcemap. Sole maintainer since 2026.04; refactored it into a TypeScript/Bun monorepo with owner packages, dependency-boundary checks and behavioral tests.
- Multi-provider agent loop: One agent loop serving Anthropic (OAuth or API key), OpenAI-compatible endpoints, ChatGPT Codex OAuth and Gemini.
- Upstream evidence pipeline: Vendored a bun-demincer fork that decodes newly installed upstream binaries, computes structural deltas across versions and has an agent write semantic reports. Behavior is ported only through reviewed, tested commits (verified through v2.1.207).
- Release engineering: Tagged releases build five standalone binaries with SHA-256 checksums, a CycloneDX SBOM and Sigstore provenance; CI boot-tests macOS, Linux and Windows, and nightly drift checks open and close an issue automatically.
- Data pipeline: Aligned balance sheets, income statements, cash flow statements, listing information and patent data under a unified firm identifier, yielding about 2,000 firm-level samples.
- Weak supervision: With no ground-truth innovation score, built 58 indicators across three dimensions (R&D, operations, support) and generated soft labels from percentile-rank scores weighted by the entropy weight method.
- Modeling: Combined financial indicators, patent technology-evolution features and BGE text embeddings; compared six models (CatBoost, XGBoost, LightGBM, ElasticNet, Lasso, MLP) with an 80/20 split, 5-fold cross-validation and ablation over three feature groups. CatBoost performed best, and patent-technology and operating features contributed most.
- Agent: A LangGraph ReAct agent with tools to score a firm, generate reports and run the pipeline, returning SHAP-based attributions; a Streamlit chat UI shows its reasoning and tool calls, and long-running jobs require confirmation.
Regional Industrial Chain Analysis and Identification of Potential Investment Targets2025.11 – 2026.03
- Industrial data processing: Processed about 400,000 business-registration records and multi-source unstructured patent data, and built upstream–downstream industry relations through semantic matching.
- Data-analysis agent: Co-developed an agent that closes the loop of data collection, Python code generation and execution, result feedback and continued analysis, so it can adjust its statistical analysis based on execution results.
- Target company screening: Identified gaps in the local industrial chain from patent-based upstream–downstream relations, and screened potential upstream and downstream companies nationwide as investment-attraction targets, supporting regional industry research and investment decisions.
Econ Atlas: Journal Crawler and Terminal Reader2025.11 – 2026.03
- Crawling: Python pipeline covering 39 top economics and management journals, including Chinese ones via CNKI, with per-publisher crawlers (Oxford, Wiley, Chicago, ScienceDirect, Cambridge, INFORMS, NBER) and enrichers that fill abstracts missing from RSS feeds.
- Reliability: Playwright with a persistent browser profile for bot-protected publishers, checkpointed resumable crawls, and a weekly macOS launchd schedule.
- Translation: Translates abstracts into Chinese with eight parallel DeepSeek workers and stores results as per-journal JSON.
- Reading: A TypeScript/Ink terminal UI with search, panel navigation and a Chinese/English abstract toggle; the Python side is typed with pyright and tested with pytest.