Chain of Title

• 

  The manifest Intake Form

• 

  The Governance Tool

• 

  The Remediation Tool

• 

  The Verification Tool

• 

  The Director Tool

Inspiration

Completion bond insurers are beginning to require proof that every AI tool used in a production was trained on licensed data with a clean chain of title - but no tooling exists to check this automatically. Productions track AI tool provenance by hand, if at all. We looked at a few angles first (VFX vendor-matching, general E&O rights clearance) before landing here specifically because it's a governance gap, not just a marketplace problem, and no dedicated tooling for it exists yet.

What it does

Chain of Title checks every AI tool listed in a production's shot manifest against a licensing registry in real time. Cleared tools pass through. Anything flagged, unverified, or tied up in active litigation gets automatically held - and the system suggests a cleared substitute in the same category where one exists. Every decision, with its reasoning and a confidence score, is published to a Kafka event stream and written to a permanent, timestamped audit log - the paper trail a completion bond insurer would actually ask to see.

How we built it

Four autonomous agents run in sequence, orchestrated by a Director built on a real Google ADK SequentialAgent - imported and run via Runner/InMemorySessionService, not just named in the README. A Verification Agent checks each tool against a Firestore-backed registry of 17 real, cited AI tools. A Remediation Agent writes holds, triggers a Cloud Function notification, and finds substitutes. A Governance Agent - built specifically using IBM Bob to satisfy our IBM partner track requirement -consumes decision events from a local Kafka (Redpanda) pipeline and logs them permanently. The Flask backend is deployed live on Google Cloud Run; the frontend is a four-page React/Vite dashboard covering Intake, Verification, Remediation, and Governance.

Challenges we ran into

Two stand out. First, a gRPC/fork() bug caused the pipeline to hang indefinitely on its first Firestore write in production - traced to Google's auth library shelling out to gcloud mid-run, fixed with GRPC_ENABLE_FORK_SUPPORT=false. Second, after building the Director as plain procedural Python early on, we caught our own architecture docs overclaiming real Agent Builder/ADK usage - so partway through the build we migrated it to a genuine ADK SequentialAgent of custom BaseAgent steps, parity-tested against the original before it replaced it in production.

Accomplishments that we're proud of

The full four-agent pipeline runs end-to-end against real Firestore and Kafka data, not mocked. The backend is genuinely deployed and verified live on Cloud Run. The Governance Agent was built with IBM Bob as a real, satisfied requirement rather than a documentation claim. The registry holds 17 real AI tools with cited, current evidence for each status, including a documented MidJourney lawsuit and the OpenAI Sora shutdown.

What we learned

Verifying requirements against primary sources - instead of pattern-matching from past experience - saves real time. Several early assumptions (an IBM watsonx MCP requirement, a Confluent Cloud dependency) turned out to be wrong and were only caught by reading the actual hackathon resource pages directly.

What's next for Chain of Title

Deploying the React frontend to Firebase Hosting so the full interactive dashboard - not just the backend API - is live at a public URL, and expanding the tool registry as more productions and AI tools are added.

Built With

• cloud-functions
• cloud-run
• docker
• firestore
• flask
• gemini
• google-adk
• google-cloud
• ibm-bob
• kafka
• python
• react
• redpanda
• vite