encyc - Overview

🤖 About Me

I build multi-agent AI systems that collaborate, debate, and execute — from trading cryptocurrency to writing long-form novels. By day, I design credit risk models; by night, I orchestrate LLM agents.

justin:
  focus: ["Multi-Agent Systems", "LLM Orchestration", "Credit Risk"]
  stack:
    agents: ["pi-agent-core", "Memori", "custom tool systems"]
    backend: ["Python", "TypeScript", "FastAPI", "Node.js"]
    data: ["ClickHouse", "PostgreSQL", "SQLite", "Faiss"]
    ml: ["XGBoost", "LightGBM", "Logistic Regression", "sentence-transformers"]
  interests:
    - Agent-native architecture (tools → agents → outcomes)
    - Multi-model routing (deep-thinking vs quick-thinking LLMs)
    - AI safety harnesses & constraint layers
  currently: "Pushing the boundary of what coordinated LLM agents can do"

🚀 Featured Projects

📈 Vibe Trading — Multi-Agent Crypto Trading

12 LLM agents. 4-phase pipeline. Real money on the line.

A production-grade cryptocurrency trading system where 12 specialized AI agents collaborate across 4 phases — Analysts (Technical, Fundamental, News, Sentiment), Researchers (Bull/Bear debate + Manager), Risk (Aggressive/Neutral/Conservative), and Decision (Trader + Portfolio Manager) — to analyze markets and execute trades on Binance.

Highlights: BM25 memory system · Dual-model routing (deep vs quick) · Prime Agent safety harness · Three-thread architecture (Macro / On-Bar / Event) · 23 structured tools · VaR + Kelly position sizing · Paper-trading testnet · Real-time web dashboard

✍️ Novel Vibe — 灵器 AI · Multi-Agent Novel Writing

1 Chief Editor + 5 specialized writing agents. Novels, not chatbots.

A structured multi-agent system where a Chief Editor orchestrates five sub-agents — Worldbuilder, Character Designer, Outliner, Chapter Director, and Chapter Writer — to co-create long-form fiction with the user. Every novel lives as portable Markdown files with YAML frontmatter in S3, indexed by PostgreSQL.

Highlights: Delete-approval protocol (sub-agents can't delete) · Per-agent Memori memory · Credit billing system · Agent Skills (agentskills.io) · Protocol bridge (craft-agents-oss ↔ custom agent server) · JWT + OAuth (Google/GitHub) · Docker + launchd + Nginx deployment · Live at app.lingqiai.art

🃏 Yihuier — 一会儿 · Credit Scorecard Toolkit

"Easily solve credit scorecard modeling — in a moment."

An object-oriented Python package that wraps the full credit scorecard pipeline: EDA, data preprocessing, ChiMerge binning, XGBoost-based variable selection, model evaluation, scorecard implementation, and PSI monitoring. Published on PyPI under MIT license.

Highlights: Full pipeline in one class · ChiMerge & custom binning · XGBoost variable selection · WOE encoding · Scorecard scaling · PSI monitoring · Built-in visualization · 80%+ test coverage

🛡️ Zhizu Super-Agent — Internal Risk Intelligence Platform

AI-powered risk control for a fintech lending platform.

An internal suite of AI/ML services powering risk decisions at 广州智租: automated credit scorecard modeling (multi-target XGBoost + LightGBM + Logistic Regression), real-time address similarity detection (sentence-transformers + Faiss vector search), gang-risk identification via phone-number graph analysis, and daily/weekly automated risk reports (Vintage, DPD, order-flow).

Highlights: Address similarity API (<100ms with cache) · Gang-risk graph detection · Multi-target model auto-selection · Automated email reporting · PMML model export · ClickHouse OLAP for billion-row risk queries · Third-party data vendor evaluation framework

⚙️ Tech Radar

📊 Stats

🏆 GitHub Trophies

🤝 Let's Connect