fix: batch inference to prevent OOM + add /ping for UptimeRobot

app.py

@@ -4,15 +4,25 @@ import torch
 
 app = Flask(__name__)
 
-# Load your specific SciBERT model from HF
+# Load model once at startup
 MODEL_PATH = "gsstec/aegis-scibert-technical"
 tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
 model = AutoModelForSequenceClassification.from_pretrained(MODEL_PATH)
+model.eval()  # disable dropout — reduces memory and is correct for inference
+
+# Process this many categories per forward pass.
+# 16 keeps peak RAM well under the free-tier limit while still being fast.
+BATCH_SIZE = 16
 
 @app.route('/')
 def index():
     return render_template('index.html')
 
+@app.route('/ping', methods=['GET'])
+def ping():
+    """UptimeRobot keep-alive endpoint — returns 200 immediately, no model inference."""
+    return jsonify({"status": "ok", "message": "TEC App is alive"}), 200
+
 @app.route('/predict', methods=['POST'])
 def predict():
     data = request.json
@@ -124,20 +134,36 @@ def predict():
         "Consciousness Transfer", "Digital Immortality", "Synthetic Life", "Artificial Evolution"
     ]
     tech_scores = {}
-    
-    for category in categories:
-        # Create category-specific input text
-        input_text = f"Scientific and technological advancements in {category} emergent in the year {year}."
-        
-        # Tokenization
-        inputs = tokenizer(input_text, return_tensors="pt", truncation=True, max_length=512)
-        
-        # Prediction
+
+    # Build all input texts upfront
+    texts = [
+        f"Scientific and technological advancements in {cat} emergent in the year {year}."
+        for cat in categories
+    ]
+
+    # Process in batches to avoid OOM on CPU-only free tier
+    for i in range(0, len(categories), BATCH_SIZE):
+        batch_cats  = categories[i : i + BATCH_SIZE]
+        batch_texts = texts[i : i + BATCH_SIZE]
+
+        # max_length=64 is plenty for these short sentences; saves ~8x memory vs 512
+        inputs = tokenizer(
+            batch_texts,
+            return_tensors="pt",
+            truncation=True,
+            max_length=64,
+            padding=True,
+        )
+
         with torch.no_grad():
             outputs = model(**inputs)
-            # Get the first prediction score for this category
-            prediction = torch.softmax(outputs.logits, dim=1).tolist()[0]
-            tech_scores[category] = prediction[0]
+            probs = torch.softmax(outputs.logits, dim=1)  # shape: (batch, num_labels)
+
+        for j, cat in enumerate(batch_cats):
+            tech_scores[cat] = float(probs[j][0])
+
+        # Free batch tensors immediately to keep peak RAM low
+        del inputs, outputs, probs
 
     return jsonify({
         "year": year,