Product thesis
NUDG treats agent access as a governed resource problem: proposal, authorization, execution, verification, and receipt are separate steps.
Pablo Zavala · AI Safety Evaluation · Research Engineering
NUDG: Governed AI Resource Operating System
NUDG is a CMU AI Venture Studio project for controlling how agents use real resources. It separates proposal, authorization, execution, verification, and audit records instead of giving agents broad access up front.
Authorization, constraints, verification, and receipts for agent-run work
Public visuals explain the system model; live product claims require separate proof packets.
Role: Founder and product builder.
How to Inspect This Work
Visual status
The detail image shows a real verification-spine preview. Concept visuals are used only where they explain the system model.
Reader check
Treat the project as founder and systems work unless a linked proof packet or live artifact is provided for a specific claim.
Case Study
Problem
Agentic tools can act faster than people can inspect them, yet most permission models still grant broad access up front.
Setup
NUDG treats resources as governed endpoints. An agent must pass through authorization, evidence, verification, and receipt layers.
Method
The product direction separates proposal, approval, execution, and audit so each step can be checked or revoked.
Result
The goal is to turn completed agent work into a portable proof object, not just a completion claim.
Limitation
Some visuals are concept artifacts. The live product work remains separate from public claims until each claim has proof.
Evidence
Public-facing materials and linked product surfaces are shown only when they are safe to publish.
Key Outcomes
- Separates proposal, authorization, execution, verification, and receipt layers
- Frames agent-run work as a portable proof object rather than a vague completion claim
- Developed through the CMU AI Venture Studio
Methods
- Agent governance
- Authorization design
- Audit records
- Product systems