Inspiration

I wanted to answer a simple question: if an AI agent uses private information, who should be allowed to see its work? TrustGraph makes source permissions and human approval part of publication.

What it does

TrustGraph tracks the inputs behind AI-generated work, carries their access restrictions into the output, and requires approval of the exact version being published. Changes to the work or its publication context require fresh review.

How we built it

I built the verification engine and gateway in Rust, used HydraDB for connected evidence, and added browser-based review tools. The coding workflow connects Codex to GitHub draft PRs. A separate 8090 integration routes GitHub issue proposals through human approval and Composio.

Challenges we ran into

The hardest parts were preserving restrictions across revisions, invalidating approval when context changes, and handling uncertain publication outcomes without accidentally sending twice.

Accomplishments that we're proud of

I built working approval and publication flows, interactive demos, and tests covering permission boundaries, stale approvals, revocation, and restart recovery.

What we learned

I learned that permission to read data does not imply permission to share work derived from it. Approval needs to cover specific content and a specific destination, backed by recorded evidence.

What's next for TrustGraph

I plan to add enterprise identity integration/authorization, broader agent-context capture, and publishing adapters

Built With

Share this project:

Updates

Submission history