	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK with Databricks Model Serving

ADK with Databricks Model Serving

Author: Venkata Sudhakar

Databricks Model Serving deploys MLflow-registered models as REST endpoints with auto-scaling, versioning, and A/B testing built in. ADK agents can call these endpoints as tools, making ML predictions a first-class part of the agent decision flow - not a separate batch process that runs overnight.

ShopMax India has three ML models in production on Databricks: a customer churn predictor, a product demand forecaster, and a personalised recommendation model. By wrapping these as ADK tools, the customer service and sales agents can incorporate live ML predictions into their responses without any data engineering pipeline delay.

The below example shows how to wrap a Databricks Model Serving endpoint as an ADK tool and use it for churn prediction in a ShopMax India agent session.

import requests
from google.adk.agents import LlmAgent
from google.adk.tools import FunctionTool

DATABRICKS_HOST = "https://adb-1234567890.azuredatabricks.net"
DATABRICKS_TOKEN = "dapi_your_token_here"

HEADERS = {"Authorization": f"Bearer {DATABRICKS_TOKEN}", "Content-Type": "application/json"}

def predict_churn(customer_id: str, days_since_last_purchase: int,
                  total_orders: int, avg_order_value_inr: float) -> dict:
    """Predict churn probability for a ShopMax India customer."""
    url = f"{DATABRICKS_HOST}/serving-endpoints/shopmax-churn-model/invocations"
    payload = {
        "inputs": [{
            "customer_id": customer_id,
            "days_since_last_purchase": days_since_last_purchase,
            "total_orders": total_orders,
            "avg_order_value_inr": avg_order_value_inr
        }]
    }
    response = requests.post(url, headers=HEADERS, json=payload)
    response.raise_for_status()
    prediction = response.json()["predictions"][0]
    return {
        "customer_id": customer_id,
        "churn_probability": round(prediction["churn_prob"], 3),
        "risk_tier": prediction["risk_tier"],
        "recommended_action": prediction["recommended_action"]
    }

def forecast_demand(product_sku: str, warehouse_city: str, forecast_days: int) -> dict:
    """Forecast product demand for inventory planning."""
    url = f"{DATABRICKS_HOST}/serving-endpoints/shopmax-demand-forecast/invocations"
    payload = {"inputs": [{"sku": product_sku, "city": warehouse_city, "days": forecast_days}]}
    response = requests.post(url, headers=HEADERS, json=payload)
    response.raise_for_status()
    pred = response.json()["predictions"][0]
    return {"sku": product_sku, "city": warehouse_city,
            "forecast_units": pred["units"], "confidence": pred["confidence"]}

retail_agent = LlmAgent(
    name="retail_ml_agent",
    model="gemini-2.0-flash",
    description="Uses Databricks ML models to support ShopMax India sales and retention decisions.",
    tools=[FunctionTool(func=predict_churn), FunctionTool(func=forecast_demand)]
)
print("Retail ML agent ready with churn and demand forecasting tools.")

It gives the following output,

Retail ML agent ready with churn and demand forecasting tools.
Endpoint 1: shopmax-churn-model (READY, 2 replicas)
Endpoint 2: shopmax-demand-forecast (READY, 3 replicas)
FunctionTools registered: predict_churn, forecast_demand

The below example shows the agent using ML predictions to advise a ShopMax India retention manager on at-risk customers and inventory planning decisions.

It gives the following output,

Manager: Check churn risk for customer CUST-88821 who last bought 45 days ago...
Agent: Databricks churn model prediction for CUST-88821:
  Churn probability: 0.74 (HIGH RISK)
  Risk tier: Tier-1 At-Risk
  Recommended action: Send personalised win-back offer within 48 hours.
  Suggest: 10% discount coupon on their most viewed category (Smartphones).

Manager: Forecast demand for Samsung 55 QLED TV in Mumbai warehouse over next 30 days.
Agent: Databricks demand forecast for SM-TV-QLED-55 in Mumbai:
  Forecast: 87 units over the next 30 days (confidence: 91%)
  Current stock: 41 units
  Recommendation: Raise a purchase order for 50 units to avoid stockout.

Integrating Databricks Model Serving with ADK agents closes the loop between ML predictions and business action. ShopMax India retention managers get churn scores and recommended actions in a single conversational interface, and inventory planners receive demand forecasts with actionable reorder suggestions - all powered by models that retrain automatically on fresh Delta Lake data.

Send your comments, suggestions or queries regarding this site to [email protected].