10. Hugging Face API deployment

We also deployed the tuned XGBoost in Hugging Face and created a Streamlit API to generate predictions for individual transactions.

10.1 Generating the app

Since I had already trained and registered the XGBoost pipeline in Databricks, I just logged the trained pipeline to the MLFlow registry and loaded it in a Hugging Face Space backed by a Git repository via MLFlow.

Specifically, We used the following app.py and Dockerfile.

Click to expand/hide Hugging Face app.py code
 ```python
import streamlit as st
import mlflow.sklearn
import pandas as pd

# Load model once (update model_uri accordingly)
model_uri = "./my_model" #"models:/workspace.default.fraud_detection_pipeline_model/1"
model = mlflow.sklearn.load_model(model_uri)

st.title("Fraud Detection Model Interface")
st.write("Enter transaction details below to get a fraud prediction:")

# Input fields matching your TransactionFeatures schema
TX_AMOUNT = st.number_input("Transaction Amount", min_value=0.0)
TX_DURING_WEEKEND = st.selectbox("Transaction During Weekend?", [0, 1])
TX_DURING_NIGHT = st.selectbox("Transaction During Night?", [0, 1])
Cust_Nb_Tx_1Day = st.number_input("Customer Number of Transactions in Last 1 Day", min_value=0)
Cust_Avg_Amt_1Day = st.number_input("Customer Avg Amount Last 1 Day", min_value=0.0)
Cust_Nb_Tx_7Day = st.number_input("Customer Number of Transactions in Last 7 Days", min_value=0)
Cust_Avg_Amt_7Day = st.number_input("Customer Avg Amount Last 7 Days", min_value=0.0)
Cust_Nb_Tx_30Day = st.number_input("Customer Number of Transactions in Last 30 Days", min_value=0)
Cust_Avg_Amt_30Day = st.number_input("Customer Avg Amount Last 30 Days", min_value=0.0)
Term_Nb_Tx_1Day = st.number_input("Terminal Number of Transactions in Last 1 Day", min_value=0)
Term_Risk_1Day = st.number_input("Terminal Risk Level Last 1 Day", min_value=0)
Term_Nb_Tx_7Day = st.number_input("Terminal Number of Transactions in Last 7 Days", min_value=0)
Term_Risk_7Day = st.number_input("Terminal Risk Level Last 7 Days", min_value=0)
Term_Nb_Tx_30Day = st.number_input("Terminal Number of Transactions in Last 30 Days", min_value=0)
Term_Risk_30Day = st.number_input("Terminal Risk Level Last 30 Days", min_value=0)

if st.button("Predict Fraud"):
    input_dict = {
        "TX_AMOUNT": TX_AMOUNT,
        "TX_DURING_WEEKEND": TX_DURING_WEEKEND,
        "TX_DURING_NIGHT": TX_DURING_NIGHT,
        "Cust_Nb_Tx_1Day": Cust_Nb_Tx_1Day,
        "Cust_Avg_Amt_1Day": Cust_Avg_Amt_1Day,
        "Cust_Nb_Tx_7Day": Cust_Nb_Tx_7Day,
        "Cust_Avg_Amt_7Day": Cust_Avg_Amt_7Day,
        "Cust_Nb_Tx_30Day": Cust_Nb_Tx_30Day,
        "Cust_Avg_Amt_30Day": Cust_Avg_Amt_30Day,
        "Term_Nb_Tx_1Day": Term_Nb_Tx_1Day,
        "Term_Risk_1Day": Term_Risk_1Day,
        "Term_Nb_Tx_7Day": Term_Nb_Tx_7Day,
        "Term_Risk_7Day": Term_Risk_7Day,
        "Term_Nb_Tx_30Day": Term_Nb_Tx_30Day,
        "Term_Risk_30Day": Term_Risk_30Day,
    }
    input_df = pd.DataFrame([input_dict])

    prediction = model.predict(input_df)[0]
    proba = model.predict_proba(input_df)[0, 1]  # Probability of class 1 (fraud)

    # st.success(f"Fraud Prediction: {'Fraudulent' if prediction == 1 else 'Legitimate'}")
    st.write(f"Probability of fraud: {proba:.2%}")

```


Click to expand/hide Hugging Face Dockerfile instructions
 ```Dockerfile
# Use official Python 3.9 base image
FROM python:3.9

# Add non-root user for permissions compliance in HF Spaces
RUN useradd -m -u 1000 user

# Set working directory
WORKDIR /app

# Copy requirements.txt separately for Docker cache efficiency
COPY --chown=user ./requirements.txt requirements.txt

# Install dependencies including streamlit and mlflow (add to your requirements.txt if missing)
RUN pip install --no-cache-dir --upgrade -r requirements.txt

# Copy all app code
COPY --chown=user . /app

# Fix permissions
RUN mkdir -p /app/mlruns && chown -R user:user /app/mlruns

# Use non-root user
USER user

# Set environment variables for user path and app caches
ENV HOME=/home/user \
    PATH=/home/user/.local/bin:$PATH \
    STREAMLIT_HOME=/tmp/.streamlit \
    HF_HOME=/tmp/huggingface \
    STREAMLIT_BROWSER_GATHER_USAGE_STATS=false

# Expose Streamlit default port
EXPOSE 7860

# Run Streamlit app, binding to all interfaces and port 7860 (HF Spaces uses this port)
CMD ["streamlit", "run", "app.py", "--server.port=7860", "--server.address=0.0.0.0"]

```


The Hugging Space app is live at: Hugging Face API

It is a simple Streamlit app where a user can enter values for the features of a transaction and select “Predict Fraud” to see the probability that a transaction with the selected features is fraudulent.

Hugging Face API_page_1 Hugging Face API_page_2


← Previous: 9. Databricks deployment Next: Post to come! →</a>