Ship-Date Drift & Inventory Buffer Orchestrator

• 

  Inbound Drift Orchestrator: Driven by Gemini 3.7 Flash

Inspiration

My own professional experience in manufacturing and distribution of hi tech and semiconductor industry, a single supplier delay email triggers hours of manual work across SAP, Excel, freight brokers, and warehouse teams while revenue exposure grows unchecked. This workflow is deterministic enough to automate yet complex enough to need intelligence - the exact sweet spot for a multi-agent AI system.

What it does

The orchestrator ingests a supplier delay notification and runs a 4-step remediation workflow in under 60 seconds: Model Armor security screening → ERP Bill-of-Materials impact analysis → parallel Warehouse + Freight sub-agent dispatch → 1-click human approval card with full credit claim and revised Sales Order dates.

How we built it

Google ADK on Vertex AI Agent Engine, Gemini 3.7 Flash, Node.js 18 tool server on Cloud Run, Cloud Firestore + SQLite persistence, Model Armor (REP) + Cloud DLP + GCP OIDC zero-trust security, Cloud Pub/Sub + Cloud Tasks for async events, OTel + Cloud Trace + Cloud Monitoring, React 18 WebSocket dashboard, Cloud Build CI/CD.

Challenges we ran into

Model Armor's global endpoint returns HTTP 403 after TOS acceptance - the regional endpoint (modelarmor.{location}.rep.googleapis.com) is the required path and is not prominently documented. VertexAiMemoryBankService requires the Agent Engine's numeric ID, which only exists post-deploy, forcing a deliberate two-pass deployment pattern.

Accomplishments that we're proud of

All 7 Google Agent Platform pillars fully implemented and validated - none mocked. A 3-layer security pipeline (Model Armor → Cloud DLP → local regex) with before_model and after_model callbacks on every LlmAgent, true parallel sub-agent execution, and zero plain-text credentials anywhere in the codebase.

What we learned

Agent topology matters more than model choice -SequentialAgent + ParallelAgent with explicit output_key state handoff made the workflow deterministic and fully unit-testable. Every GCP integration being fire-and-forget (fail-open) means the system degrades gracefully in dev without real cloud credentials.

What's next for Ship-Date Drift & Inventory Buffer Orchestrator

Live SAP S/4HANA OData connector, Slack/Teams 1-click approval cards, a negotiation sub-agent to autonomously engage alternate suppliers when SLA thresholds are breached, and predictive drift detection using time-series models on Firestore/BigQuery delivery history.

Built With

• adk
• agent
• ai
• engine
• google
• vertex