FlowOps AI

An agentic AI workforce that processes business documents across markets using Gemini, Google ADK, and Cloud Run, with local rules, human review, and full auditability.

Comment

Inspiration

Business document processing is still highly manual, fragmented, and difficult to scale across countries. FlowOps AI was created to demonstrate how an agentic AI workforce can transform a business document into a validated, auditable operational decision while respecting country-specific rules.

What it does

FlowOps AI is a multi-country agentic document operations platform.

A document enters the system and is processed through an orchestrated workflow that can:

  • Extract structured business data
  • Detect the document's country
  • Apply country-aware validation rules
  • Make an operational decision
  • Route exceptions to Human Review
  • Detect duplicate documents
  • Send approved records to a Mock ERP
  • Generate a complete audit trail

The current implementation supports Brazil and the United States.

How we built it

FlowOps AI uses Google Agent Development Kit (ADK) to orchestrate specialized agents including DocumentAgent, ValidationAgent, DecisionAgent, and ReportingAgent.

Gemini 3.6 Flash performs structured AI extraction, with a deterministic local parser available as a fallback.

The backend is built with FastAPI.

Google Cloud infrastructure

FlowOps AI's backend is deployed and running on Google Cloud Run, fulfilling the project's Google Cloud infrastructure requirement.

Deployment:

  • Google Cloud project: flowops-ai-506712
  • Cloud Run service: flowops-ai
  • Region: southamerica-east1
  • Backend: FastAPI + Google ADK
  • AI model: Gemini 3.6 Flash
  • Secret management: Google Secret Manager
  • Public backend: https://flowops-ai-vns7icztma-rj.a.run.app

The demo video provides direct visual proof of the live deployment through the real Google Cloud Console, showing the Cloud Run service and the running .run.app backend.

Google technologies used

  • Google Agent Development Kit (ADK)
  • Google GenAI SDK
  • Gemini 3.6 Flash
  • Google Cloud Run
  • Google Secret Manager

Challenges

One of the main challenges was creating an agentic workflow that remained auditable and deterministic enough for business operations while still benefiting from generative AI.

Another challenge was supporting documents from different countries without forcing every market into the same validation rules.

We also designed the system so Gemini failures do not automatically break the workflow: a local parser fallback and Human Review path provide additional operational resilience.

What we learned

We learned that production-oriented agentic systems need more than an LLM call. They need orchestration, deterministic validation, fallback paths, human escalation, auditability, and clear infrastructure boundaries.

We also learned how Google ADK, Gemini, and Google Cloud Run can work together as different layers of the same agentic architecture.

What's next

The next evolution of FlowOps AI is focused on production persistence, authentication and tenant isolation, real document intake channels, controlled pilot testing, and gradual integration with enterprise systems.

The hackathon version intentionally uses a Mock ERP and keeps future production integrations separate from the demonstrated implementation.

Built With

Share this project:

Updates

Submission history