Perpetual

Inspiration

Research does not fail because scientists cannot chat with a model. It stalls because literature search, hypothesis formation, preregistration, experiments, critique, provenance, and reporting are fragmented across tools. Most assistants end when the chat ends.

What it does

Perpetual gives an AI/ML team a persistent Research Room led by Ava. Ava can search governed scholarly sources, propose falsifiable hypotheses, ask focused clarification questions, seal plans before execution, run bounded asynchronous missions, challenge results, and leave a reproducible Research Pack.

Five Google ADK specialists cover literature, methods, criticism, verification, and writing. Human turns, autonomous runs, citations, tool receipts, and confirmed memory stay in one inspectable timeline. Work survives reloads and controlled restarts, while ambiguous or sensitive decisions pause in a Needs You queue instead of being guessed.

How we built it

The Next.js workbench connects through a same-origin BFF to a FastAPI control plane. Gemini 3.7 Flash is the only live model route, reached through Google ADK and the Google GenAI SDK on Vertex AI. MCP tools sit behind exact allowlists, typed schemas, risk labels, timeouts, idempotency fences, and payload-free receipts.

The verified Google Cloud demo runs the backend and campaign surface on Cloud Run. Cloud Scheduler advances a bounded campaign every five minutes using a dedicated OIDC service account plus a rotated Secret Manager wake token. The deployed revision passed credentialed Gemini 3.7 Flash structured-output and tool-calling canaries, and the successful Cloud Build source inputs byte-match the candidate's Docker inputs.

After capturing the deployment proof, we locked Cloud Run to the owner and the dedicated Scheduler identity. Anonymous health, readiness, and campaign requests now return HTTP 403; authenticated owner checks and the OIDC Scheduler wake return HTTP 200. Public-access prevention is enforced on both Cloud Storage buckets, and there are no public project, repository, bucket, or secret principals.

AI output is not accepted as scientific truth by itself. Plans are registered before execution, tools return provenance, Critic and verifier paths are separate, promotion requires independent recomputation, and the final manuscript is bound to a content-addressed Research Pack with a rerun recipe.

Why it matters

An agent becomes a collaborator when it preserves the researcher's intent and the work's lineage—not when it produces more prose. Perpetual keeps bounded research moving between conversations and overnight while ensuring people decide what matters and evidence decides what survives.

Built for Collaborative Partner

Perpetual targets the Collaborative Partner category: it works alongside a researcher over time, maintains shared context safely through compaction and durable state, delegates to specialists, uses tools transparently, and brings the human back exactly when judgment is required.

Evidence and reproducibility

Source and reproducible setup: https://github.com/ychampion/perpetual Google Cloud proof: https://github.com/ychampion/perpetual/blob/main/docs/hackathon-build/google-cloud-proof.md