Vice CEO: Proof-Carrying Business Autonomy
Vice CEO is a behind-the-scenes business operator that prepares routine customer communication and approved follow-through, carries the evidence behind every decision, and brings a person in only when judgment is actually needed.
The problem
Small business teams lose time to work that is individually simple but collectively relentless: support questions, sales follow-ups, policy lookups, changing public program information, and the handoff between an alert and the next useful action. Typical AI tools can generate a reply, but they leave the business to reconstruct why it acted, whether it had authority, and what happened afterward.
The solution
Vice CEO turns a bounded business signal into prepared, inspectable work. It connects the source, applicable policy, proposed action, authority boundary, and delivery state in one replayable receipt. Routine customer-service responses and consent-aware outreach follow-ups are prepared in the background; exceptions such as a refund request return as a focused human decision.
Westover EPR is the example deployment. Its private Registry Change Watch observes an approved public-source portfolio, preserves hash-linked evidence in Firestore, detects material changes, and prepares an owner-facing operational brief. The public reviewer app uses closed synthetic fixtures so judges can inspect the entire workflow without customer data or external business effects.
How AI is used
- A four-role Google Agent Development Kit specialist fleet is configured around the locked
gemini-3.5-flashtarget. - Specialists are capability-limited: Support Intake can read only a named synthetic case; policy, escalation, and routing roles have no direct business-action tools.
- A separately deployed fixed-prompt Gemini connectivity canary produced a hash-only Cloud Logging receipt with no customer data, tool calls, persistent business writes, or external business effect.
- The private Registry Change Watch uses bounded Gemini briefs only after an approved public source materially changes. The local judge flow stays deterministic and provider-free so it is reproducible.
Key features
- Prepared customer work — policy-linked customer-service replies and consent-aware outreach follow-ups, each with a zero-effect public-demo receipt.
- Human judgment escalation — exceptions are surfaced with the relevant evidence instead of being silently actioned.
- Proof-carrying decisions — source, policy, decision, authority, and delivery state stay linked and inspectable.
- One-use Action Warrant — a signed, short-lived, role-scoped warrant authorizes one synthetic transition and deterministically rejects a replay or tampering attempt.
- Background Registry Change Watch — a private Cloud Scheduler and Cloud Run worker monitors an approved public EPR source portfolio, deduplicates runs, and stores hash-linked Firestore evidence.
- Fail-closed controls — unknown inputs, stale or replayed warrants, scope violations, changed controls, and untrusted source content are rejected or escalated.
- Reproducible reviewer proof — a public synthetic walkthrough, proof bundle, evidence replay, and automated adversarial/evaluation checks let reviewers inspect the implementation without credentials.
Built with
Gemini 3.5 Flash, Google Agent Development Kit, Cloud Run, Firestore, FastAPI, and Python.
Findings and learnings
Useful autonomy needs evidence and boundaries, not just fluent output. Vice CEO separates recommendation, authority, delivery, and proof so an organization can start with preparation and expand execution only for workflows it explicitly approves. The public reviewer stays synthetic and reproducible, while the private operational worker is independently bounded to approved public sources and Firestore evidence state.
The public demo uses synthetic business cases so reviewers can inspect the workflow without exposing customer data or sending real customer or prospect messages.
Built With
- fastapi
- gemini-3.5-flash
- google-agent-development-kit-(adk)
- google-cloud-firestore
- google-cloud-run
- python

Log in or sign up for Devpost to join the conversation.