app.py

import onnxruntime as ort
import gradio as gr
import numpy as np

# ------------------------------------------------
# Load ONNX model (CPU only)
# ------------------------------------------------
sess = ort.InferenceSession(
    "collusion_xgb_model.onnx",
    providers=["CPUExecutionProvider"]
)

# ------------------------------------------------
# Detect probability output tensor safely
# ------------------------------------------------
output_info = sess.get_outputs()
PROB_INDEX = None

for i, out in enumerate(output_info):
    if len(out.shape) == 2 and out.shape[1] == 2:
        PROB_INDEX = i
        break

if PROB_INDEX is None:
    raise RuntimeError("❌ Probability output not found in ONNX model")

print(f"✅ Using output index {PROB_INDEX} as probability tensor")

# ------------------------------------------------
# Risk bucket logic (PRODUCT FRIENDLY)
# ------------------------------------------------
def risk_bucket(p):
    if p >= 0.05:
        return "HIGH RISK"
    elif p >= 0.01:
        return "MEDIUM RISK"
    else:
        return "LOW RISK"

# ------------------------------------------------
# Prediction function
# ------------------------------------------------
def predict(amount, user_txn, driver_txn, pair_count, hour, day):
    """
    Feature order MUST match training:
    [amount, user_txn_count, driver_txn_count, user_driver_pair_count, hour, day_of_week]
    """

    X = np.array(
        [[
            float(amount),
            float(user_txn),
            float(driver_txn),
            float(pair_count),
            float(hour),
            float(day)
        ]],
        dtype=np.float32
    )

    outputs = sess.run(None, {sess.get_inputs()[0].name: X})
    probs = outputs[PROB_INDEX]

    fraud_prob = float(probs[0][1])  # class-1 probability

    # Convert to interpretable risk
    risk = risk_bucket(fraud_prob)
    risk_score = int(fraud_prob * 1000)  # scaled score for visibility

    return fraud_prob, risk_score, risk

# ------------------------------------------------
# Gradio UI
# ------------------------------------------------
gr.Interface(
    fn=predict,
    inputs=[
        gr.Number(label="Amount"),
        gr.Number(label="User Txn Count"),
        gr.Number(label="Driver Txn Count"),
        gr.Number(label="User–Driver Pair Count"),
        gr.Number(label="Hour (0–23)"),
        gr.Number(label="Day of Week (0=Mon, 6=Sun)"),
    ],
    outputs=[
        gr.Number(label="Fraud Probability"),
        gr.Number(label="Risk Score (0–1000)"),
        gr.Text(label="Risk Level"),
    ],
    title="Collusion Fraud Detection (ONNX)",
    description=(
        "This model detects potential user–driver collusion. "
        "Fraud probability is a ranking signal; risk score and bucket are used for decisions."
    ),
).launch()