Create app.py

app.py

@@ -0,0 +1,145 @@
+import gradio as gr
+import spaces
+import torch
+import json
+import re
+import urllib.request
+import os
+from transformers import AutoProcessor, AutoModelForMultimodalLM
+
+MODEL_ID = "google/gemma-4-E2B-it"
+
+print(f"Loading {MODEL_ID}...")
+processor = AutoProcessor.from_pretrained(MODEL_ID)
+model = AutoModelForMultimodalLM.from_pretrained(
+    MODEL_ID,
+    torch_dtype=torch.bfloat16,
+    device_map="auto"
+)
+print("Model loaded successfully.")
+
+os.makedirs("sample_data", exist_ok=True)
+SAMPLE_IMAGE = "sample_data/car_damage.jpg"
+SAMPLE_AUDIO = "sample_data/driver_statement.wav"
+
+if not os.path.exists(SAMPLE_IMAGE):
+    urllib.request.urlretrieve(
+        "https://www.driving.org/wp-content/uploads/2023/11/driver-hand-examining-dented-car-with-damaged-fend-2023-07-17-20-53-56-utc-e1699944140557.jpg",
+        SAMPLE_IMAGE
+    )
+
+if not os.path.exists(SAMPLE_AUDIO):
+    urllib.request.urlretrieve(
+        "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/apps/sample-data/journal1.wav",
+        SAMPLE_AUDIO
+    )
+
+@spaces.GPU
+def process_insurance_claim(image_path, audio_path):
+    if not image_path or not audio_path:
+        return {"error": "Both an image of the damage and an audio statement are required."}
+
+    system_prompt = """You are an expert AI Auto Insurance Claim Adjuster. 
+Your task is to analyze the provided image of vehicle damage and the audio statement from the driver. 
+Cross-reference the audio description with the visual evidence.
+You must output ONLY a valid JSON object. Do not include markdown formatting like ```json.
+The JSON must strictly follow this schema:
+{
+  "damage_severity": "Low|Medium|High|Total Loss",
+  "affected_parts": ["list", "of", "damaged", "car", "parts"],
+  "driver_statement_summary": "Short 1-sentence summary of the audio transcript",
+  "consistency_check": "Match|Mismatch",
+  "flagged_for_review": true|false,
+  "reasoning": "Brief explanation of why it matches or doesn't match the visual evidence."
+}"""
+
+    messages = [
+        {
+            "role": "system", 
+            "content": [
+                {"type": "text", "text": system_prompt}
+            ]
+        },
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "url": image_path},
+                {"type": "audio", "audio": audio_path},
+                {"type": "text", "text": "Analyze this insurance claim and output the JSON report."}
+            ]
+        }
+    ]
+
+    inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        return_dict=True,
+        return_tensors="pt",
+        add_generation_prompt=True,
+        enable_thinking=False
+    ).to(model.device)
+    
+    input_len = inputs["input_ids"].shape[-1]
+
+    outputs = model.generate(
+        **inputs, 
+        max_new_tokens=512,
+        temperature=0.2,
+        top_p=0.95,
+        top_k=64
+    )
+    
+    response = processor.decode(outputs[0][input_len:], skip_special_tokens=True)
+    
+    clean_response = re.sub(r"^```(?:json)?\s*", "", response).strip()
+    clean_response = re.sub(r"\s*```$", "", clean_response).strip()
+    
+    try:
+        json_output = json.loads(clean_response)
+        return json_output
+    except json.JSONDecodeError:
+        return {
+            "error": "Failed to parse JSON output.",
+            "raw_output": response
+        }
+
+css = """
+#component-0 { max-width: 900px; margin: auto; }
+.gr-button { background-color: #2563eb !important; color: white !important; }
+"""
+
+with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
+    gr.Markdown(
+        """
+        # 🚗 AI Auto Claim Adjuster (Gemma 4 E2B)
+        Upload a photo of the vehicle damage alongside an audio statement from the driver describing the incident. 
+        Gemma 4 E2B natively processes **both the audio wave and the image** simultaneously, transcribing the story, analyzing the visual damage, and outputting a structured JSON claim adjustment report.
+        """
+    )
+    
+    with gr.Row():
+        with gr.Column(scale=1):
+            img_input = gr.Image(type="filepath", label="1. Upload Vehicle Damage Image")
+            audio_input = gr.Audio(type="filepath", label="2. Upload Driver Audio Statement")
+            submit_btn = gr.Button("Generate Claim Report", size="lg")
+            
+        with gr.Column(scale=1):
+            json_output = gr.JSON(label="Structured Claim JSON Output")
+
+    submit_btn.click(
+        fn=process_insurance_claim,
+        inputs=[img_input, audio_input],
+        outputs=[json_output]
+    )
+    
+    gr.Examples(
+        examples=[[SAMPLE_IMAGE, SAMPLE_AUDIO]],
+        inputs=[img_input, audio_input],
+        outputs=[json_output],
+        fn=process_insurance_claim,
+        cache_examples=False,
+        label="Try Demo Example"
+    )
+
+if __name__ == "__main__":
+    demo.launch()