About

Hi, I'm Jet Xu.

Turning private work files and code into reliable, evidence-backed context for AI.

Systems Architect | 15+ years building mission-critical infrastructure | AI Harnessing & Context Engineering

Reasoning is improving fast. Reliable context is still the bottleneck.

Now: building DocMason, a local-first, evidence-first knowledge base for AI-assisted deep research over private work files.
Before: built code-intelligence systems across llama-github, LlamaPReview, and repo-graph-rag.
Direction: the Mason ecosystem—moving from deep document analysis to multi-stage pipelines that generate editable, consulting-grade native deliverables.

Why This Path

Over 15 years of architecting mission-critical systems, the recurring failure mode is always the same: in high-stakes environments, being “almost right” is useless. The bottleneck to reliable output—whether from humans or AI—is rarely raw reasoning capacity. It is context precision. That constraint drives everything I build.

Current Focus

DocMason is my current open-source focus: a local-first, provenance-first knowledge base for AI-assisted deep research over private work files. It is not a document chatbot. It compiles unstructured artifacts into knowledge infrastructure that agents can actually use. Its native operating pattern is simple: the repo is the app, and Codex is the runtime.

Core architectural priorities:

Deterministic ingestion: parsing PDFs, decks, spreadsheets, emails, and repository-native text without silent failures.
Reliable outputs: provenance-first retrieval instead of vague, hallucination-prone document chat.
Actionable output: extending the Mason ecosystem beyond extraction. The next step is a deterministic, multi-stage pipeline separating semantic narratives from visual layout DOMs, turning deep document analysis directly into editable, consulting-grade native presentations (PPTX) for serious white-collar work.

The Foundation

I came to document intelligence through code intelligence.

llama-github: the retrieval substrate, built to give LLMs GitHub-native context instead of raw repository dumps.
LlamaPReview: field validation for that thesis. It achieved a measured 61% signal-to-noise ratio in AI code review across 4,000+ active repositories (35K+ combined stars).
repo-graph-rag: the Code Mesh research artifact, exploring deterministic repository graphs and traversal-first retrieval.
llamapreview-context-research: formalizing the exact failure mode of Context Instability.

This path started with helping AI understand code diffs, but led to a broader conclusion: the real computing frontier is shifting toward understanding full knowledge environments and generating high-stakes output from them. Code Mesh was the logical end of one inquiry, but not the final product surface.

The Pivot

By late 2025, it became clear that reactive code review would not remain the terminal surface of AI engineering. As autonomous agents and “vibe coding” accelerated, the scarce problem was no longer commenting on diffs after the fact, but making sure the right constraints, context, and operating surfaces exist before generation happens.

That shift is why I chose not to commercialize my earlier graph-based intelligence explorations, and instead focused on proactive context infrastructure and deterministic output pipelines.