Professor of Medical AI @ SYSU

Building clinically grounded AI, Codex-first stage-led agent infrastructure, the One Person Lab desktop workbench, independent domain agents, and auditable workflows for high-value knowledge work.

This profile serves as a public entry point for my academic and technical work in medical AI, research systems, and clinically grounded workflow design.

The current public surfaces are:

• fenggaolab.org for academic information
• One Person Lab as the Codex-first, stage-led OPL Framework for automated delivery of high-value knowledge work
• One Person Lab App as the desktop workbench, installer, release, and user-facing product repository for OPL
• Med Auto Science as an independent medical research domain agent with a single med-autoscience app skill
• Med Auto Grant as an independent medical grant domain agent with a single med-autogrant app skill
• RedCube AI as an independent visual-deliverable domain agent with a single redcube-ai app skill

One Person Lab is the OPL Framework: a complete agent runtime framework for a one-person research lab and other high-value knowledge work. It is Codex-first and stage-led: large tasks are organized into stages that resemble how human experts execute the work, and Codex CLI is the smallest execution unit inside each stage.

It currently owns:

• the framework-level stage model, session/runtime coordination, progress projection, and artifact discovery surfaces
• Codex-first execution through CLI/MCP/product-entry contracts, app skills, and explicit domain-agent activation
• shared family contracts, indexes, provider adapters, health checks, and thin coordination surfaces
• support for external runtime providers where useful, without moving domain truth out of the owning domain agents
• machine-readable interfaces consumed by the desktop App and other shells

One Person Lab App is the separate user-facing desktop product. It owns the installer, desktop packages, release assets, updater metadata, first-run checks, screenshots, and user documentation. It packages the OPL Framework, domain agents, and companion tools into a workbench for daily use, while the framework repo keeps runtime, activation, contracts, module management, and machine-readable App interfaces.

Domain agents keep their own canonical domain truth, stable callable surfaces, audit writeback, runtime records, and delivery boundaries.

The public product model is:

OPL Framework -> One Person Lab App -> Foundry Agents

Human
  -> One Person Lab App for daily workbench use
      -> One Person Lab framework
          -> Codex-first session/runtime
          -> Stage-led task execution
          -> Progress, files, settings, health, and artifact discovery
          -> Explicit domain-agent activation when specialized work is needed
              -> Research Foundry -> Med Auto Science app skill
              -> Grant Foundry -> Med Auto Grant app skill
              -> Presentation Foundry -> redcube-ai app skill

Developer / operator path
  -> One Person Lab framework
      -> Codex-first session/runtime
      -> Stage-led task execution
      -> Progress, files, settings, health, and artifact discovery
      -> Explicit domain-agent activation when specialized work is needed
          -> Research Foundry -> Med Auto Science app skill
          -> Grant Foundry -> Med Auto Grant app skill
          -> Presentation Foundry -> redcube-ai app skill

Execution unit:
  Codex CLI inside a stage

Integration/reference layer:
  CLI, MCP, product-entry contracts, runtime provider adapters, OPL Runtime Manager projections, App-consumed interfaces, and repo-tracked schemas

MedDeepScientist is no longer a MAS default runtime dependency. It remains only as a MAS-declared source-provenance, historical-fixture, explicit archive-import, backend-audit, upstream-intake, and parity-oracle reference, not as a top-level public domain agent.

Future workstreams such as IP Foundry, Award Foundry, Thesis Foundry, and Review Foundry stay in definition until their domain package and public boundary are ready.

Med Auto Science is the current active medical research implementation on the Research Foundry line.

It focuses on medical research operations, including:

• disease-workspace organization
• data governance
• study progression
• evidence packaging
• manuscript and submission delivery

Med Auto Grant is the active medical grant implementation on the Grant Foundry line.

It focuses on author-side grant operations, including:

• direction screening
• question refinement
• argument building
• proposal drafting
• critique, revision, and package export

RedCube AI is the active visual-deliverable implementation on the Presentation Foundry line.

It currently focuses on:

• ppt_deck and presentation-grade visual delivery
• visual deliverable governance and review loops
• agent-first production of audited visual artifacts

Distribution-wise, the App repository owns desktop packaging, releases, updater metadata, and user-facing installation. The domain agents are consumed through OPL-managed module installation and current Packages/GHCR-backed coordinates, while each domain repo keeps its own authority over domain truth and delivery readiness.

Together, the current public shape is:

One Person Lab -> explicit MAS activation -> Med Auto Science

alongside:

One Person Lab -> explicit MAG activation -> Med Auto Grant

alongside:

One Person Lab -> explicit RCA activation -> RedCube AI

My current work is centered on:

• clinically grounded AI for colorectal cancer
• agent-first, human-auditable domain systems
• medical research workflow design
• Codex-first, stage-led agent infrastructure for high-value knowledge work

• Academic site: fenggaolab.org
• One Person Lab: gaofeng21cn/one-person-lab
• One Person Lab App: gaofeng21cn/one-person-lab-app
• Med Auto Science: gaofeng21cn/med-autoscience
• Med Auto Grant: gaofeng21cn/med-autogrant
• RedCube AI: gaofeng21cn/redcube-ai