Curatom

Inspiration

Every time a different model is asked to act for a business, the business's own facts get reinvented. Agents leak across tenants. Policy lives in a doc, not on the request path. Curatom exists so a company has one canonical record of what it is and what it wants — and so every agent that acts on it is registered, authorized, and auditable.

This is Comfort Curators Private Limited's Fortified Enterprise Fleet entry for Google Cloud's All Things Agentic hackathon.

Pre-existing work (required disclosure)

Curatom is prototype initially started on 26.08.26. rv0.2.0 is on Zenodo (https://zenodo.org/records/22112980): the tenant-scoped registry, policy engine, grounded recall, residency enforcement, and control-plane UI.

All of it is Built during the Submission Period (3–31 Aug 2026):

Google ADK 2.8.0 fleet on Gemini 3.5 Flash via Vertex AI (ADC, no API key)
Durable runtime: Cloud Tasks POST /tasks → /tasks/execute
Live Cloud Run proof at GET /ops/gcp-proof
First-party Model Armor equivalent (prompt-injection / tool-poisoning / PII / residency 403s)
A2A-shaped Agent Cards and reasoning-chain traces
Grounded Decision Log writes from the fleet — the agent takes an action, not only a reply

What it does

A judge registers their own workspace. That tenant is empty until they add context, an atom, and a memory. Then they run a fleet goal. The ADK agent recalls only what that tenant is allowed to see, writes a decision with citations, and refuses jailbreaks and residency mismatches instead of returning an empty list.

How we built it

One Cloud Run service serves FastAPI and the built React UI. Firestore holds documents and 768-d gemini-embedding-001 vectors. Secret Manager holds JWT and task secrets. Vertex AI is keyless from the runtime service account. SequentialAgent 429'd this project's quota; the live fleet is one ADK Agent with gateway / memory / orchestrator tools. We wrote that down rather than pretend otherwise.

Challenges

Vertex quota killed SequentialAgent. Cloud Run CPU idles between requests, so Cloud Tasks is a real HTTP callback, not a background thread. Google Model Armor is a separate product we did not buy — the track accepts first-party equivalents, and ours is a heuristic.