Engineer & Quant Trader · Founder of DealMesh · Kaggle Gold (9/3,856) · NeurIPS 2026 Agenthon T3 Prize (3/55) · JoinQuant research Sharpe 2.499
工程师 & 量化交易员 · DealMesh 创始人 · Kaggle 金牌（9/3,856）· NeurIPS 2026 Agenthon T3 获奖（3/55）· 聚宽研究回测 Sharpe 2.499

• 🥇 Kaggle team gold — 9th / 3,856 (top 0.23%) · Home Credit — Credit Risk Model Stability · certificate issued to X. Dxfory · Kaggle 团队金牌
• 🏆 NeurIPS 2026 Competition Track — Agenthon T3 Prize, 3rd / 55 · Team abc123 · T2 8th/87 · T4 26th/63 · T1 48th/61 · T3 获奖、四赛道参赛
• 📈 Huatai Securities factor research — Rank IC 0.080–0.100 · long–short Sharpe up to 3.23 · CSI 1000 · internship research backtest · 华泰中证 1000 多因子实习研究回测
• 📊 JoinQuant research backtest — Sharpe 2.499 · return 70.90% · annualized 73.56% · excess 37.00% · max drawdown 20.11% · benchmark 399101.XSHE return 24.74% · alpha 0.521 · beta 0.810 · ¥100,000 · 2024-11-15 → 2025-11-15 · 聚宽研究回测
• ⚡ fastsim — 880× median same-machine speed-up · row-exact Rust rewrite of ABIDES · Team abc123 · Rust 市场仿真
• 🔌 edgar-mcp — shipped on PyPI + MCP Registry · SEC filings for AI agents, every number cited
• 🚀 DealMesh — Private beta · public sign-up opening soon · 452 commits · 3,286 passing tests · Trust Receipts for PE/VC research & due diligence · 内测中 · 即将向公网开放注册
• 🧪 QBX — hundreds of platform-certified strategies across US + A-share, L1/L2 data · 美股 + A 股数百条平台认证策略
  • Currently contributing to QBX's ongoing platform refinement and financial-agent development. · 目前参与 QBX 的后续完善与金融 agent 构建。

I build evidence-grounded AI systems for finance — agents, models, and research infrastructure from the first pipeline to a shipped product. 我做有出处的金融 AI：智能体、模型与研究基础设施，从第一条数据管道做到可交付产品。

Private beta · public sign-up opening soon 内测中 · 即将向公网开放注册

DealMesh — Trust Receipts for PE/VC research and due diligence. Built 0→1 since May 2026: 452 commits, ~210,000 lines of Python, ~46,000 lines of TypeScript, 369 test files, 3,286 passing tests. Demand validation through 20+ PE/VC institution interviews. Huatai Chuangxing Global CEO Accelerator.

The product · 产品： DealMesh attaches a verifiable Trust Receipt to every research conclusion — pinned to the source passage and graded L0–L5. DealMesh 为每条投研结论生成可核验的 Trust Receipt：定位原文，并按 L0–L5 分级证据充分度。

The problem · 痛点： In due diligence, the question that decides the work is whether a conclusion is supported by the source passage — the question I worked through on ITDD deals at EY. 尽调里决定工作质量的问题，是这条结论能否由原文段落撑住——我在安永 ITDD 项目里直接面对过这个问题。

• 🛠️ Built 0→1: data pipeline → evidence binding → export gate → signed receipts → web app → cloud deploy
• ⚙️ Engineering scale · 工程规模: 452 commits since May 2026 · ~210,000 lines of Python · ~46,000 lines of TypeScript · 369 test files · 3,286 passing tests
• 🤝 Demand validation · 需求验证: 20+ PE/VC institution interviews in China and abroad
• 🌏 Markets: Mainland China · Hong Kong · Singapore
• 🏆 Selected into Huatai Chuangxing Global CEO Accelerator · 入选华泰创星全球 CEO 加速营

Architecture · 架构

flowchart LR
 A["Deal materials<br/>PDF, Excel, filings"] --> B["Ingest and parse"]
 B --> C["Evidence binding<br/>source passage pin"]
 C --> D["Trust Receipt<br/>L0-L5"]
 D --> E["Export gate<br/>human review"]
 E --> F["Signed receipt bundle"]
 F --> G["Web app / cloud"]

DealMesh product UI · demo diligence case. Chat and judgment cards are the product surface. MedPilot / 智医科技 in the screenshot are demo project names.
DealMesh 产品界面 · 演示尽调案例。 对话与判断卡片是产品界面。截图中的 MedPilot / 智医科技是演示项目名。

Team abc123 poster, NeurIPS 2026 Agenthon Workshop, Atlanta, 12 December 2026. T3 Prize, 3rd of 55.

• Published MCP · 已发布 MCP — edgar-mcp: SEC 10-K/10-Q segment revenue, BDC non-accrual, Form 4 — every figure with accession and an EDGAR URL (PyPI 1.1.3 + MCP Registry)
• Competition agents · 竞赛智能体 — Team abc123 at Agenthon 2026 (public repo), NeurIPS 2026 Competition Track: T3 Prize 3rd/55, T2 8th/87, T4 26th/63, T1 48th/61, and a row-exact Rust market simulator (fastsim)
• Parallel LLM collection · 并行采集 — strategy-factory campaigns run 4 parallel LLM collectors, configured ceiling 6
• Multi-model routing · 多模型路由 — DealMesh routes 11 model tasks with call logging, usage accounting, and an 8M-token project budget cap
• Evaluation gates · 评测闸门 — evidence-match set of 514 cases (494 scored); release gate requires Cohen's κ ≥ 0.75 on human-adjudicated labels
• Full-stack 0→1 · 全栈 — FastAPI · Celery · PostgreSQL/pgvector · Redis · React/Vite · Docker · Alibaba Cloud

• Huatai Securities (华泰证券) — CSI 1000 multi-factor research: Rank IC 0.080–0.100, long–short Sharpe 2.57–3.23 (internship research backtest)
• QBX strategy system — US / HK / A-shares × L1 / L2 / L3. Hundreds of platform-certified strategies across US and A-share markets, on L1/L2 data. 1,700+ backtests run through the research pipeline across the QBX platform and an in-house strategy factory. 690+ structured strategy IRs from the same research pipeline. Engine alignment Sharpe 1.0814 vs 1.0820; Deflated Sharpe ≥ 0.95 gates; IC corr 0.998 · qbx
• A-share five-model ensemble — JoinQuant research backtest, benchmark 399101.XSHE, ¥100,000, 2024-11-15 → 2025-11-15. v2: Sharpe 2.499, return 70.90%, annualized 73.56%, excess 37.00%, max drawdown 20.11%, alpha 0.521, beta 0.810. v1: Sharpe 2.033, return 58.49%, annualized 60.69%, excess 27.05%, max drawdown 23.49% · 5machine_learning
• Crypto network science — temporal motifs on USDT / USDC / DAI / UST / LUNA transfer graphs, including the May 2022 window. Lead figures: USDT_1h_motifs_heatmap.png and motif_heatmap_null_model.png
• Competitions — Kaggle team gold 9th/3,856 · Agenthon Team abc123 T3 Prize 3rd/55, T2 8th/87, T4 26th/63, T1 48th/61 · 2026 Hong Kong Quant Trading Championship, live-trading track, active participant

作品目录 · 证据对应表

📫 Evidence-grounded AI agents · quant research · due diligence