About

About

Hi there! 👋

I’m Jet Xu, an engineer passionate about building AI systems that understand code at scale. My work focuses on Repository Graph RAG architectures, LLM context engineering, and evidence-based code intelligence — making AI agents truly comprehend software repositories through structured knowledge graphs.

I believe the future of automated development lies in Code Graph RAG — where AI agents reason about codebases through persistent knowledge graphs capturing classes, methods, dependencies, and call relationships. That’s the core of what I’m building.

🎯 Core Research & Engineering Focus

Repository Graph RAG Architectures

Building next-generation retrieval systems specifically for code understanding:

  • Code Knowledge Graphs capturing structural relationships (classes, methods, imports, call graphs)
  • Hybrid retrieval pipelines combining graph traversal, semantic search, and LLM-powered reasoning
  • Multi-source context assembly from self repositories, public codebases, issues, and discussions
  • Production-grade graph databases (ArangoDB) for efficient traversal and complex query patterns

LLM Context Engineering for Automated Development

Designing intelligent context retrieval that powers AI-driven development workflows:

  • Evidence-based reasoning grounding AI insights in actual code snippets and architectural context
  • Multi-stage retrieval with embedding models, rerankers, and confidence scoring
  • Task-aware context selection adapting retrieval strategies based on development task complexity
  • Repository pool caching for real-time performance in multi-agent systems

Privacy-Preserving Code Intelligence

Building secure architectures for enterprise code analysis:

  • Zero-knowledge storage using cryptographic blind indexing (HMAC-SHA256)
  • Ephemeral processing where code content never persists, only graph structures
  • Federated learning approaches for cross-repository insights without data sharing

🚀 Featured Projects

LlamaPReview — Evidence-Based AI Code Review

A GitHub App delivering Graph RAG-powered PR reviews with context retrieval and severity gating for low-noise, high-impact insights.

Core Architecture:

  • 🧠 Context Retrieval Engine: Finds related, unchanged code to surface ripple effects early
  • 📊 Evidence-Backed Findings: Every high-priority issue ties to real code snippets with confidence scores
  • 🎯 Severity Gating: Deterministic filtering reduces false positives and review fatigue
  • 🔒 Secure Knowledge Graph: Professional tier uses ArangoDB for persistent code structure analysis

Technical Innovations:

  • Repository Graph RAG: Classes, methods, dependencies, and call graphs enable deep semantic understanding
  • Zero-Knowledge Architecture: Code content is ephemeral; only cryptographic fingerprints persist
  • Self-Adaptive Review: Learns from your team’s review patterns over time
  • Multimodal Analysis: Inline suggestions, architectural diagrams (Mermaid), and structured summaries

Current Adoption:

  • 🌟 3,000+ active repositories (32,000+ combined stars)
  • 🆓 Free for public repos with automatic advanced analysis on complex PRs
  • 🚀 Professional tier (launching soon) with full knowledge graph capabilities
Try LlamaPReview →Website →

llama-github — Foundation for Automated AI-Driven Development

An open-source Python library for efficient GitHub knowledge retrieval, designed as the context retrieval module for future automated development systems powered by multi-agent LLMs.

Vision: Automated AI-Driven Development

The library serves as a foundational component in a larger architecture (see diagram above) that enables:

  • Multi-Agent Orchestration: LLM-based agents collaborating on complex coding tasks
  • Retrieval-Augmented Development Loop: Continuous context retrieval from self repos and public knowledge
  • Task & State Management: Intelligent task allocation, scheduling, and state tracking
  • Development Tools Integration: File editing, testing, version control, and deployment automation

Current Capabilities:

  • 📚 Dual-Mode Retrieval: Self repositories + public GitHub resources (code, issues, discussions)
  • 🎯 LLM-Powered Query Analysis: Intelligent search strategy generation
  • 🔄 Advanced Ranking: Embedding similarity + reranker scores + LLM relevance scoring
  • ⚡ Production-Ready: Async processing, repository pool caching, flexible authentication
  • 🗄️ Vector Database Integration: Efficient storage and retrieval at scale

Quick Start:

1

pip install llama-github

1
2
3
4
5
6
7
8
9
10

from llama_github import GithubRAG

github_rag = GithubRAG(
    github_access_token="your_token",
    openai_api_key="your_key"
)

# Retrieve context for development tasks
query = "How to implement async context managers in Python?"
context = github_rag.retrieve_context(query)

Future Development: While currently focused on LlamaPReview development, llama-github will evolve to support the full automated development vision — enabling AI agents to autonomously accomplish complex coding tasks through intelligent context retrieval and tool orchestration.

License: Apache 2.0Status: Available on PyPI

View on GitHub →

🛠️ Technical Expertise

AI/ML & Code Intelligence

Python LangChain Hugging Face

  • Repository Graph RAG: ArangoDB for code knowledge graphs, graph traversal algorithms
  • LLM APIs & Models: OpenAI GPT-4o, Anthropic Claude, Google Gemini, Mistral AI
  • Embedding & Reranking: Jina AI embeddings and rerankers, sentence-transformers
  • Code Analysis: AST parsing, static analysis, tree-sitter, semantic understanding
  • Vector Databases: Efficient storage and retrieval for embedding-based search

Enterprise Architecture & Cloud-Native Systems

Java Docker

  • Cloud-Native Architecture: Microservices, containerization, auto-scaling, high availability
  • Cloud Platforms: Alibaba Cloud, AWS (compute, storage, serverless)
  • High-Concurrency Systems: Distributed architectures, message queues, caching strategies
  • Domain-Driven Design: Modular architecture, service boundaries, API design

Search & Data Technologies

  • Semantic Search: Multimodal search (text, image, voice), embedding-based retrieval
  • Graph Databases: ArangoDB for complex relationship modeling and traversal
  • Databases: MySQL, MongoDB, Redis, distributed data systems
  • Data Platforms: Customer data platforms, real-time processing, analytics

📝 What I Write About

On this blog, you’ll find deep dives into:

  • Repository Graph RAG: Architectures for building code knowledge graphs that power AI agents
  • LLM Context Engineering: Strategies for intelligent context retrieval in automated development
  • Code Intelligence Systems: Building AI that reasons about software architecture, not just syntax
  • Privacy-Preserving AI: Cryptographic techniques for secure code analysis without persistent storage
  • Production ML Systems: Lessons learned from deploying AI tools at scale (3,000+ repos)
  • Multi-Agent Orchestration: Designing systems where LLM-based agents collaborate on complex tasks

I also share insights from enterprise architecture, digital transformation, and the intersection of AI with real-world software engineering challenges.

🌟 Current Initiatives

Short-Term (Q4 2025)

  • 🚀 Launching LlamaPReview Professional with ArangoDB-powered knowledge graphs for private repos
  • 📊 Publishing benchmarks: Code Graph RAG vs. vector-only RAG for software understanding tasks
  • 📝 Writing comprehensive guides on Repository Graph RAG architectures

Long-Term Vision: Automated AI-Driven Development

Building towards fully automated development systems (as illustrated in the architecture diagram) that combine:

  • Multi-Agent Orchestration: LLM-based agents collaborating through intelligent task allocation
  • Repository Graph RAG: Persistent code knowledge graphs enabling deep architectural reasoning
  • Development Tools Integration: Seamless file editing, testing, version control, and deployment
  • Continuous Learning: Self-adaptive systems that improve from developer feedback
  • Privacy-First Design: Zero-knowledge architectures for enterprise adoption

The vision: AI agents that can autonomously accomplish complex coding tasks by intelligently retrieving context from self repositories and public knowledge, while maintaining human oversight through conversation management.

💡 Philosophy

“The best AI tools are evidence-based — they ground insights in real code structures, not hallucinations.”

I believe in building AI systems that:

  • ✅ Reason through graphs — Code understanding requires structural knowledge, not just embeddings
  • ✅ Ground in evidence — Every insight ties back to actual code snippets and relationships
  • ✅ Prioritize signal over noise — Severity gating and confidence scoring reduce false positives
  • ✅ Respect privacy — Cryptographic blind indexing, zero-knowledge storage, ephemeral processing
  • ✅ Learn continuously — Adapt to team patterns, improve through usage

🎓 Background

With 15+ years spanning enterprise architecture and AI/ML engineering, I bring a unique perspective to building practical AI systems:

  • AI/ML Engineering: Hands-on development of LLM-powered applications, Graph RAG systems, and AI agents
  • Enterprise Architecture: Designing mission-critical platforms (CRM, CDP, e-commerce) serving millions of users
  • Technical Leadership: Leading cross-functional teams (architecture, DevOps, QA) in multinational environments
  • Open Source Contribution: Building developer tools (llama-github, LlamaPReview) used by the global community

My experience includes architecting cloud-native platforms with auto-scaling and high availability, implementing semantic search engines with multimodal capabilities (image, voice), and designing privacy-compliant data systems for regulated industries.

I’ve also led digital transformation initiatives from 0 to 1, including:

  • China Unified Services Platform at Kering (luxury retail’s strategic business mid-platform)
  • Digital Mid-Platform at Volvo Cars (20+ microservices, 10+ system integrations)
  • Social Network Mobile App as founder/CEO (millions of users, ML-powered personalization)

This combination allows me to bridge cutting-edge AI research and production-ready systems that developers actually want to use.

🏆 Recognition & Achievements

  • 🥇 First Award at National Olympiad in Informatics in Provinces (NOIP)
  • ⭐ Top Contributor (Top 5%) at IBM for three consecutive years (2012-2014)
  • 🚀 3,000+ repositories actively using LlamaPReview
  • 📦 Open-source maintainer of llama-github (Apache 2.0 license)

📫 Get in Touch

I’m always interested in discussing:

  • Repository Graph RAG architectures for code intelligence
  • Challenges in building production LLM systems with privacy guarantees
  • Multi-agent orchestration for automated development
  • Collaboration opportunities on open-source AI tooling
  • Feedback on LlamaPReview and llama-github

Connect with me:

📚 Acknowledgments

My work builds on the shoulders of giants. Special thanks to:

  • ArangoDB for the powerful multi-model database enabling complex graph queries
  • Jina AI for excellent embedding models and reranker APIs
  • LangChain for the foundational LLM application framework
  • The open-source community for countless libraries, tools, and shared knowledge

🔖 Fun Facts

  • 🦙 The “Llama” in my project names honors Meta’s Llama models and the open-source LLM movement
  • 🎓 Double bachelor’s degree in Biological Sciences & Computer Sciences from Fudan University
  • 🎯 My goal: Build AI agents that can autonomously accomplish complex coding tasks through intelligent context retrieval
  • 📖 Firm believer that code understanding requires graphs, not just embeddings

Last Updated: October 2025

This page is a living document. Check back for updates on Repository Graph RAG research and automated development progress!