AI Agent Architect

Jerry Liu Portfolio

Co-founder and CEO of LlamaIndex

Co-founder and CEO of LlamaIndex, the data framework used to connect LLM agents to external knowledge sources; his GitHub profile shows the underlying "llamaagents" project series — scoped agent architectures for tasks like arXiv summarization, patent search, and financial statement extraction, built on LlamaIndex's retrieval and orchestration primitives.

RAGAgent OrchestrationLlamaIndexOpen Source

Location

Location: classified. Send a pigeon.

Companies

LlamaIndex

Skills

RAGAgent OrchestrationLlamaIndexOpen Source

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

A general framework demonstrated through many small, scoped agents

Rather than one flagship demo, the llamaagents-* repositories each apply the same orchestration primitives to a distinct real task (patents, court orders, resumes), showing the underlying architecture is genuinely reusable, not built for a single showcase.

2

Framework maintenance visible alongside application work

The profile shows both the core llama-hub data-loader library and applied agent projects side by side, demonstrating the same person doing infrastructure and application work rather than only one layer of the stack.

3

Forks of adjacent frameworks kept visible

Maintaining visible forks of DSPy and CrewAI alongside his own framework signals active engagement with how competing approaches to agent orchestration solve the same problems, not a closed-off, single-framework view.

What AI Agent Architects can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • Demonstrate a general framework through several small, scoped example agents rather than one flagship demo — it proves the architecture is reusable.

  • Show both the infrastructure you maintain and the applications built on it in the same place — it demonstrates full-stack ownership of the pattern.

  • Keep visible forks or experiments with competing frameworks in your space — it signals you track the field, not just your own corner of it.

Projects

The framework is the well-known part and the recent repos say more. The creator of LlamaIndex still ships small, runnable document-agent apps under his own name (form filling, classification, parsing), and each one is a working demo of where the company is heading.

Open SourceadvancedOngoing since Nov 2022

LlamaIndex

PythonRAGAgentsVector storesLlamaParseLiteParse

An open-source framework for building LLM applications over your own data. It started with RAG and grew into agents and workflows. You install a core package plus integrations for your chosen LLM, embedding model and vector store. The README now says the project's focus has moved to document parsing and extraction (LlamaParse, LiteParse).

Keeping a 300-plus-integration ecosystem coherent is a lot of architecture work. The README also explains in plain terms why the company changed direction.

  • Core + integration packages under a namespaced import scheme

  • 300+ integrations for LLMs, embeddings and vector stores

  • Repo now centers on document parsing: LlamaParse, LiteParse, ParseBench

Web AppintermediateCommit history Dec 2025 – Jan 2026

PDF Form Filler

Claude Agent SDKMCPFastAPINext.js 16React 19PyMuPDFLlamaParseServer-Sent Events

A full-stack app that fills PDF forms from natural-language instructions. A FastAPI backend runs the Claude Agent SDK with custom MCP tools. PyMuPDF detects and edits form fields, and LlamaParse pulls context from uploaded PDFs, DOCX, PPTX or images.

A complete agent app rather than a notebook: tool definitions, document parsing, streaming UI and session state all connected.

  • Agent edits real form fields through custom MCP tools

  • Upload context files; the agent extracts the values it needs

  • Streaming Next.js 16 UI with original vs filled PDF views

ToolintermediateCreated Sep 2026

DocJev

PythonCLILiteParseJevLlamaParse (optional OCR)uv

A Python library, CLI and local app that classifies a PDF, DOCX or PPTX into categories, or splits a packet into its component documents, using category rules written in plain language. LiteParse extracts page text locally and Jev predicts the category or the document boundaries, with probabilities and review flags.

It turns a common enterprise chore (splitting and routing document packets) into a measurable tool, with separate OCR and decision timings. Letting a coding agent maintain the repo is itself a live experiment in agent workflows.

  • Classify: one document → category, probabilities, review flags

  • Split: one packet → ordered documents with page ranges and optional per-document PDF export

  • Issue triage, reviews and merges handled by Codex, per the README

Ready to build your portfolio?

Follow the AI Agent Architect roadmap — skills, projects, and timeline to get hired.

AI Agent Architect Roadmap