ProcureFlow

Inspiration

Procurement teams often compare supplier proposals in spreadsheets while important facts remain scattered across documents. ProcureFlow turns that handoff into a clear, evidence-backed decision workflow.

What it does

A buyer fixes the request and budget first. ProcureFlow preserves three synthetic proposal records, compares only verifiable facts, and sends a request-bound selection task to Gemini 3.5 Flash through Vertex AI. Gemini proposes; it cannot approve, contact suppliers, or change commercial facts. A separate synthetic reviewer checks the proposal before the decision is completed.

The hosted demo shows a completed office-laptop procurement with the recommendation, supporting evidence, independent review, and audit history. It uses synthetic data only: no supplier, email, phone, payment instrument, or external business system is contacted.

How it was built

Google ADK runs the bounded agent step inside a Process Engine that owns authorization, workflow state, idempotency, and protected effects. The public Cloud Run web service calls an IAM-private API. Cloud SQL stores workflow state, Memorystore supports rate limiting, and Cloud Storage preserves immutable synthetic evidence. The repository includes the architecture diagram and local setup instructions.

What I learned

The central challenge was keeping the AI useful without making it an authority. The model returns only a bounded request-local selection; the application derives references, evidence, and risk signals from immutable facts. A distinct reviewer is required for completion, and invalid output fails closed to a safe fallback.

Built for this hackathon

This project was created for the All Things Agentic Hackathon.

Built With

• artifact-registry
• cloud-run
• cloud-sql
• cloud-storage
• docker
• gemini-3.5-flash
• google-adk
• google-genai
• memorystore
• postgresql
• python
• secret-manager
• typescript
• vertex-ai