ai-research-os

Building Your Own AI Research OS Workshop

This is the repo for the talk at the AI Engineer World's Fair conference,
presented by Louis-François Bouchard (X · LinkedIn) and Paul Iusztin (X · LinkedIn).

---

I'm always losing my research.

5,000+ notes in Obsidian, 5,000+ highlights in Readwise, plus Notion and Google Drive — growing ~250 files a month. And every research session still starts from zero: paste the same links into Codex, watch it rebuild context on the fly, then lose all of it when the chat ends.

Access to information was never the bottleneck. Reusing it was.

AI Research OS turns your Second Brain into a living research memory your agents maintain — research that compounds over weeks, months, and years instead of dying in a conversation.

Codex or Claude gives you an answer. This gives your harness a reusable research memory:

• raw/ - copied or fetched source material
• wiki/sources/ - per-source summaries
• wiki/concepts/, wiki/entities/, wiki/comparisons/ - reusable synthesis pages
• wiki/overview.md and wiki/synthesis.md - the current thesis
• index.yaml and index.md - the catalog future agents read first
• log.md and wiki/open-questions.md - what happened and what to research next

The point is not to replace Codex. The point is to stop re-researching the same topic every week.

What this is

AI Research OS is a set of local AI skills for building and querying a persistent research wiki from your own sources:

• Obsidian notes
• Readwise highlights
• NotebookLM notebooks
• GitHub repos
• YouTube videos with public transcripts
• web links
• PDFs and local files

Via deep research on top of your personal Second Brain

Obsidian is optional. It is just a visual IDE for browsing the generated markdown wiki. The system can run purely through Codex or Claude Code from a normal working directory.

Slides and video

• Slides: Google Slide Deck
• Video: [] (coming soon)

Whenever You're Ready, Here's How to Go Deeper

Our AI Research OS runs locally via skills.

Our Agent AI Engineering course, built with Towards AI, shows how to ship it to production as a multi-agent system: MCP servers with LangGraph, an evaluator-optimizer loop, observability, evals, and GCP deployment.

35 lessons. 3 end-to-end portfolio projects. A certificate. And a Discord community with direct access to industry experts and me.

Built for software engineers, data engineers, or scientists transitioning into AI engineering.

Rated 5/5 by 300+ students. The first 7 lessons are free:

Start here →

Architecture

Sources flow through deep research, get stored as raw files, indexed, synthesized into a wiki, and then queried:

user question / sources
        |
        v
  /research router
        |
        +--> query existing wiki
        +--> append known sources
        +--> run deep discovery
        |
        v
 raw sources -> source pages -> concepts/entities/comparisons
        |
        v
 index.yaml + overview.md + synthesis.md + open-questions.md

Examples

Three end-to-end runs, each browsable in examples/. Open the linked prompt in Claude Code / Codex with /research to reproduce it; each screenshot shows the resulting wiki browsed in Obsidian.

1. Deep research from an outline + web resources

Discover sources, summarize, and synthesize them into a topic wiki.