Automated Science: AI-Enhanced Research Infrastructure

Status: Active Development

Automated Science is my vision for the future of scientific research—where AI systems work alongside human scientists to accelerate discovery. This isn’t about replacing researchers, but about creating a symbiosis between human expertise, experimental knowledge, and AI’s ability to process vast amounts of information.

The Challenge

Traditional approaches to AI-led science face fundamental barriers:

1. The Authority Problem

Academics won’t publish in platforms not recognized by authoritative bodies—papers don’t count toward promotions or grants if the publication venue isn’t established.

2. The Grounding Problem

Science is tightly linked to experiments, and experiments require real humans. A lot of knowledge in laboratories is tacit—not written down, but held in the minds of experienced researchers.

3. The Integration Problem

AI systems typically operate in isolation, disconnected from the day-to-day workflow of active research groups.

The Solution: Enterprise AI-Archive

The path forward is integrating AI deeply within real research institutions:

┌─────────────────────────────────────────────────────────────┐
│                    AI-Enhanced Science Center               │
│                                                             │
│   ┌─────────────┐    ┌─────────────┐    ┌─────────────┐   │
│   │  Real Labs  │───▶│  AI Agents  │───▶│  Authority  │   │
│   │ Experiments │    │  Grounded   │    │  in Domain  │   │
│   │ Tacit Know. │    │  in Reality │    │  Expertise  │   │
│   └─────────────┘    └─────────────┘    └─────────────┘   │
│                                                             │
└─────────────────────────────────────────────────────────────┘

By embedding AI infrastructure within research centers:

• AI agents gain access to real experimental data and laboratory protocols
• Researchers get AI assistance integrated into their existing workflow
• Over time, the AI develops deep domain expertise grounded in actual practice
• Eventually, this expertise becomes authoritative for reviewing work in that domain

The Dual-Pipe Architecture

Working with research institutions requires careful handling of both code/papers and large data files. The architecture separates:

The “Brain” (Git Monorepo)

• Handles logic, code, and papers
• Strict governance via Pull Requests
• Permission-based access for different researchers

The “Body” (Syncthing/Tailscale)

• Handles heavy data and outputs
• Automated background synchronization
• Real-time sync to researcher machines

The AI Agent Bridge

services:
  agent-runtime:
    volumes:
      # Code & Papers (Read/Write via PR)
      - ./workspace:/app/workspace
      
      # Raw Data (Read-Only)
      - /srv/data/raw:/app/data/raw:ro
      
      # Results (Read-Write for outputs)
      - /srv/data/outputs:/app/data/outputs:rw

The agent reads from the “Brain” to understand the science and mounts the “Body” to execute analysis—all without bypassing safety checks for code changes.

Key Components

Cloud Tasks

Researchers submit specific questions via structured YAML files. AI processes queries through scheduled pipelines, delivering results as merge requests for human review.

Collaboration Sessions

Interactive research via browser-based VS Code connected to AI agents. Researchers can pair-program with AI, exploring data together with full access to analysis tools.

Autonomous Research

Daily scheduled jobs for autonomous data exploration. AI agents generate hypotheses, find interesting patterns, and produce research artifacts for human scientists to validate.

Current Development

This vision is being developed in collaboration with ELSC (Edmond and Lily Safra Center for Brain Sciences), where I completed my PhD. The goal is to make ELSC a leading center in collaboration between real scientists and AI.

Near-term Goals

• Deploy AI infrastructure integrated with lab workflows
• Create AI agents with deep knowledge of computational neuroscience
• Build tools for AI-assisted paper writing and peer review

Long-term Vision

Develop an agent so deeply integrated with ELSC’s labs that it becomes authoritative in reviewing papers in computational neuroscience—grounded by its understanding of real experiments, not just text.

The Bigger Picture

This approach represents an “Enterprise tier” of AI-Archive—not just a platform for AI papers, but a complete solution for making science centers AI-enhanced. Each integrated center becomes a node in a larger network where AI agents, grounded in real experimental knowledge, can meaningfully contribute to scientific progress.

Interested in bringing AI-enhanced research to your institution? Let’s talk about how to integrate AI infrastructure with your research workflow.