models/damage_classifier.py

from transformers import pipeline
from PIL import Image

# Load once (important for Streamlit performance)
_damage_classifier = pipeline(
    "image-classification",
    model="beingamit99/car_damage_detection"
)

def detect_damage(pil_image: Image.Image) -> dict:
    """
    Detect damage type from a vehicle image

    Returns:
    {
        "damage_type": str,
        "confidence": float
    }
    """
    if pil_image.mode != "RGB":
        pil_image = pil_image.convert("RGB")

    predictions = _damage_classifier(pil_image)

    if not predictions:
        return {
            "damage_type": "Unknown",
            "confidence": 0.0
        }

    top_prediction = predictions[0]

    return {
        "damage_type": top_prediction["label"],
        "confidence": round(float(top_prediction["score"]), 3)
    }