Skip to content

UC Open 2026 — A community-owned AI model for research tools

A 15-minute talk at the UC Open 2026 summit in Berkeley, April 22–23, 2026. Condensed from the long-form OSA platform overview.

Interactive Slides

Use arrow keys to navigate. Press F for fullscreen, ? for shortcuts.

Abstract

Research tools like BIDS, EEGLAB, MNE-Python, and HED depend on a handful of maintainers, scattered documentation, and forums that are easy to miss. General-purpose AI assistants hallucinate tool-specific answers. Each community is too small to build a bespoke AI on its own.

The Open Science Collective (OSC) treats this as a shared-infrastructure problem. The Open Science Assistant (OSA) is the AI layer of that infrastructure: one platform, one FastAPI + LangGraph agent loop, one SQLite FTS5 store per community, and a YAML file that onboards the next community in an afternoon. Seven assistants are live today; EEGLAB alone has answered over sixteen hundred questions at a 99% success rate, peaking above 100 questions a day.

The talk covers:

  • Why the OSC model works for small research communities that cannot each fund their own AI stack
  • The friction on both sides: stagnant communities with rotting docs, and general AI that hallucinates
  • Architecture that avoids the usual Retrieval-Augmented Generation (RAG) failure modes: source of truth is the live community (GitHub, docs, papers, mailing lists), re-synced on every PR, issue, or doc update, with citations by construction and Anthropic prompt caching for ~90% cost savings on system and repeated prompts
  • YAML-driven onboarding: a single file creates the assistant, API routes, embeddable widget, and knowledge sync job
  • Public status at status.osc.earth/osa: aggregate + per-community metrics, sync health, and usage charts
  • How to join: file an issue, open the Discord, or drop a YAML

Interactive slides built with Agentic Presentation Builder, a JSON-to-Reveal.js presentation tool.