Liangjun (Lance) Song, PhD

Forward-Deployed AI Engineer building production agentic workflows, RAG systems, and automation platforms.

View on GitHub

GitHub Project Portfolio

This page summarises the GitHub projects I have built or meaningfully contributed to. I group them by the engineering problems they demonstrate: search and retrieval systems, applied ML and AI products, agentic developer tooling, data science workflows, and human-facing AI applications.

My background is a PhD in search and recommendation systems, three years as a Data Scientist building search ranking, recommendation, and content-understanding pipelines at Redbubble (100 M+ user events), and current work on production LLM and agentic systems. These projects span that full arc — from classical retrieval and ML to modern RAG, agents, and evaluation engineering.

Applied AI Products

Project Role Stack What it demonstrates
Termly Creator / lead developer FastAPI, Pydantic, Claude Sonnet, Azure Document Intelligence, Tesseract OCR, Docker, Azure/Vercel/Supabase Australian medical contract automation: MPPA PDF generation, Medicare/PBS identifier validation, scanned PDF extraction, OCR correction, structured clause/entity extraction, and compliance-oriented risk analysis.
Video-RAG Builder / collaborator Python, CLIP, ChromaDB, OpenCV, LangChain, OpenAI/Anthropic/Gemini, uv Multimodal retrieval system for video and image collections: frame sampling, vision-language embeddings, persistent vector search, timestamped retrieval, and optional LLM answer generation.
RageZone Creator / lead developer WeChat Mini Program, Node.js, WebSocket, Firebase/JWT, OpenAI/Gemini/Deepgram, Supabase Realtime communication assistant for emotionally charged conversations, with persona/mode setup, chat UI, WebSocket service design, corpus/evaluation scripts, and product hooks for paid usage.
Elder Companion Creator / prototype Expo, React Native, Firebase, ElevenLabs, Claude, audio APIs Mobile voice companion prototype with elder-facing chat, family dashboard, message storage, daily summary planning, and alert workflow design.

Agentic Developer Tooling

Project Role Stack What it demonstrates
Remote Agent Workbench Creator / lead developer React, Vite, TypeScript, Express, Socket.IO, Cloudflare Tunnel, Codex CLI, Claude Code VS Code-like remote workbench for running Claude Code and Codex from a browser, including live terminal output, task persistence, file previews, permission prompts, model controls, and safe allowed-directory execution.
AgentForge Creator / lead developer React, Node.js, Express, Socket.IO, SQLite, xterm.js, WSL, git worktrees, PM2 Local control plane for multiple coding agents, with real-time terminals, task queues, worktree isolation, auto-commit scheduling, session logs, and Cloudflare Tunnel remote access.
Claude Code Source Map Research / analysis Node.js, source maps, CLI architecture notes Early technical study of Claude Code internals and agent CLI behavior, used to understand terminal-agent architecture, command flow, and developer workflow design.

Operator and Workflow Systems

Project Role Stack What it demonstrates
Toy Gifting System Creator / lead developer Python, SQLite, JavaScript, Three.js, supplier APIs, data-import scripts Back-office tool for toy sourcing and gift-box assembly, combining supplier catalog ingestion, compliance metadata, operator feedback, bilingual product data, and 3D packing visualization.
BranchFlow / AI Writer Creator / lead developer React, TypeScript, Vite, Tailwind CSS, Framer Motion, React Flow Local nonlinear writing and prompt orchestration workbench for Seedance-style video generation, with branching narrative state, prompt drawers, mind-map view, and Obsidian Canvas export.
Dispatch.AI Course Project Instructor / curriculum builder LangGraph, Pydantic, Redis, MCP, guardrails, RAGAS AI engineering course project where students build a booking assistant with durable state, tool routing, retrieval, guardrails, and evaluation rather than a single prompt demo.

Research Forks and Labs

Project Role Stack What it demonstrates
AI Hedge Fund Lab Research fork / study Python, multi-agent workflows, market data, backtesting Study of multi-agent financial reasoning patterns across valuation, fundamentals, sentiment, technical analysis, risk, and portfolio decision agents.
TradingAgents Research fork / study Python, LLM agents, financial workflow simulation Comparative study of trading-agent orchestration patterns and debate-style analyst roles for complex decision workflows.
Everything Claude Code Contribution / tooling study Shell, TypeScript, Python, hooks, commands, skills Agent-harness performance and workflow system studied and adapted for my own Claude Code / Codex operating practice; I treat this as tooling research, not an original product claim.

What These Projects Show