LIANGJUN (LANCE) SONG, PHD

SENIOR ML ENGINEER / DATA SCIENTIST · SEARCH & RECOMMENDATION

WORK EXPERIENCE

WiseTech Global
Apr 2025 - Present
AI Engineer - AI/ML Group
  • Architected and optimized end-to-end LLM systems for production support platforms (CWBot & TriageAgent), combining LangGraph orchestration, FastAPI services, Azure OpenAI, enterprise APIs, and durable PostgreSQL-backed state.
  • Designed RAG workflows with dynamic question rewriting, custom retriever factories, Azure Search integration, and heterogeneous internal data sources to improve how users find and act on operational knowledge.
  • Implemented guardrails for PII filtering and response steering, plus checkpointing and recovery patterns for reliable multi-turn product experiences.
  • Led container-first CI/CD modernization for AI services, establishing Docker dev containers, GitHub Actions pipelines, readiness checks, and environment parity across local and deployed workflows.
  • Debugged production-like infrastructure issues across Azure Search schemas, devcontainer networking, PostgreSQL SSL, and service configuration drift.
  • Maintained test coverage across backend, frontend, and system test sets, incorporating linting, type checks, quality gates, and release-readiness checks.
SGLang Framework
Feb 2025 - Present
Open Source Contributor
  • Contributing core optimizations to the SGLang framework (LMSYS / UC Berkeley) to accelerate LLM inference pipelines for state-of-the-art models including DeepSeek R1, Llama 3, and Qwen.
  • Working on backend runtime, distributed serving, and frontend prompt-language features to improve throughput, contextual routing, and controllability.
  • Leveraging AI coding agents (Claude Code) to navigate the large codebase, draft PRs, and review diffs, demonstrating an effective human-in-the-loop open source contribution workflow.
Australia IT Group · AI Engineer Bootcamp
Feb 2026 - May 2026
Guest Instructor & Curriculum Designer — RAG, Multi-Agent Systems, Fine-tuning
  • Delivered 12+ sessions on production RAG, embeddings & vector search, multi-agent systems (LangGraph, AutoGen, CrewAI), QLoRA fine-tuning, RAGAS evaluation, and applied AI (UI, PDF parsing, Text-to-SQL); all notebooks and slides committed to the cohort repo.
  • Designed Dispatch.AI, a 6-week hands-on course project (13 students): students build a production AI booking assistant incrementally — Pydantic state + Redis persistence, LangGraph agents, MCP tool server, multi-agent routing, NeMo Guardrails, RAGAS evaluation, and Docker + Render capstone; delivered reference implementation, scaffold, and CI/CD grading pipeline.
Redbubble
Mar 2023 - Jan 2025
Data Scientist - Search & Recommendation Team
  • Led search and recommendation enhancements with Marqo vector search and GCP Vertex AI MLOps pipelines, improving add-to-cart rate by 0.5% and CTR by 10%.
  • Owned production-shaped ML workflows from analysis and offline evaluation through experiment design, deployment coordination, and post-launch metric review.
  • Designed ML/data infrastructure over 100M+ user events for feature extraction, search relevance analysis, and downstream serving workflows.
  • Drove GA4 analytics migration to ensure data reliability across internal dashboard-driven A/B testing frameworks.
  • Optimized search relevance for long-tail queries, balancing retrieval quality, user behavior signals, and measurable business impact.
Redbubble
Aug 2022 - Mar 2023
Data Scientist - AI Renovation Team
  • Applied language-image models for automated content tagging and classification across a large marketplace catalog.
  • Enhanced content taxonomy using structured user engagement signals and measurable search/discovery outcomes.
  • Ran experiments for new content tags and search-engine optimization, translating model output into product-facing discovery improvements.
  • Developed personalization workflows that connected data science experiments with production product metrics.
Redbubble
Jan 2021 - Jul 2022
Data Scientist - Content Moderation Team
  • Implemented production content-classification systems using language-image models for automated moderation.
  • Developed image duplicate-detection pipelines using ML and statistical modeling techniques.
  • Established data quality and anomaly-detection workflows for safer model and pipeline operation.
  • Shipped classification workflows that reduced intellectual-property moderation risk and improved operational throughput.
LIANGJUN (LANCE) SONG, PHD PAGE 2

SELECTED ML / AI SYSTEMS

Termly (GitHub)
2026
Creator & Lead Developer
  • Built a document AI system for Australian medical contract automation: scanned PDF extraction, OCR correction, structured clause/entity extraction, validation, and risk analysis.
  • Implemented FastAPI/Pydantic services integrating Claude Sonnet, Azure Document Intelligence, Tesseract OCR, Docker, and healthcare-specific validation logic.
Video-RAG (GitHub collaboration)
2025
Builder & Collaborator
  • Built a multimodal retrieval pipeline that samples video frames, generates CLIP embeddings, persists them in ChromaDB, and answers natural-language queries with timestamped visual evidence.
  • Supported image-only search paths and multiple LLM providers for local testing, retrieval experiments, and provider comparison.
Agentic Workbench (Remote Agent Workbench + AgentForge · GitHub)
2025 - Present
Creator & Lead Developer
  • Built Remote Agent Workbench: browser orchestration for Claude Code and Codex with React/Vite, Express, Socket.IO, live PTY terminal streaming, task persistence, file previews, and allowed-directory safety boundaries.
  • Built AgentForge: multi-agent control plane for running Claude Code, Codex, and Kimi Code in parallel across isolated git worktrees, with xterm.js terminals, PM2 management, auto-commit scheduling, and session logs.
Toy Gifting System (github.com/lycanlancelot/toy-gifting)
2025
Creator & Lead Developer
  • Built a back-office tool for toy sourcing and gift-box assembly with SQLite catalog ingestion, supplier/compliance metadata, bilingual product data, operator feedback, and Three.js packing visualization.
  • Modeled practical sourcing constraints including small-batch suppliers, certifications, dimensions, price conversion, and operator review loops.
BranchFlow / RageZone / Elder Companion
2025 - 2026
Creator / Prototype Builder
  • Built BranchFlow, a React/TypeScript nonlinear writing and prompt orchestration workbench with branching story state, prompt drawers, React Flow mind maps, and Obsidian Canvas export.
  • Built human-facing AI prototypes including RageZone (WeChat Mini Program + Node/WebSocket communication assistant) and Elder Companion (Expo voice app with family dashboard and summary workflow).

RESEARCH EXPERIENCE

Microsoft Research Asia
Jul 2012 - Jun 2013
Research Intern - Web Search and Data Mining Group
  • Developed Autosub, a collaborative project for generating video subtitles using advanced speech-to-text APIs.
  • Contributed to web page mining projects using data mining and time-series analysis techniques.

PUBLICATIONS

Towards Efficient Personalized Ranking
PhD Thesis, RMIT University
2020
Incremental Preference Adjustment: a Graph Theoretical Approach
The VLDB Journal (Core Rank A*)
2020
Continuous Summarization over Microblog Threads
DASFAA (Best Student Paper Award Runner Up)
2017