Early Morning Improvements

services/feature_engineering.py

@@ -204,30 +204,53 @@ class SafetyFeatureEngineer:
 
         Uses percentile normalization to preserve meaningful gradients
         even when the absolute density range is narrow.
+
+        Points outside India's geographic bounds receive a low-neutral
+        default since the KDE model is trained exclusively on Indian data.
         """
         if self.kde_model is None:
             return np.zeros(len(midpoints))
 
+        # India bounding-box guard — the KDE has no training data outside
+        # these bounds, so international points get a flat neutral value.
+        INDIA_LAT_MIN, INDIA_LAT_MAX = 6.0, 37.0
+        INDIA_LON_MIN, INDIA_LON_MAX = 68.0, 98.0
+
+        in_india = (
+            (midpoints[:, 0] >= INDIA_LAT_MIN)
+            & (midpoints[:, 0] <= INDIA_LAT_MAX)
+            & (midpoints[:, 1] >= INDIA_LON_MIN)
+            & (midpoints[:, 1] <= INDIA_LON_MAX)
+        )
+
+        density = np.full(len(midpoints), 0.1)  # neutral for international
+
+        india_mask = np.where(in_india)[0]
+        if len(india_mask) == 0:
+            return density
+
         # score_samples returns log-density; exponentiate for raw likelihood
-        log_density = self.kde_model.score_samples(midpoints)
-        density = np.exp(log_density)
+        india_points = midpoints[india_mask]
+        log_density = self.kde_model.score_samples(india_points)
+        india_density = np.exp(log_density)
 
         # Apply regional crime multiplier from district/state data
-        density = density * self.regional_crime_multiplier
+        india_density = india_density * self.regional_crime_multiplier
 
         # Percentile normalization for robust [0, 1] mapping
-        if len(density) > 10:
-            p5, p95 = np.percentile(density, [5, 95])
+        if len(india_density) > 10:
+            p5, p95 = np.percentile(india_density, [5, 95])
             if p95 > p5:
-                density = np.clip((density - p5) / (p95 - p5), 0.0, 1.0)
+                india_density = np.clip((india_density - p5) / (p95 - p5), 0.0, 1.0)
             else:
-                d_max = density.max()
-                density = density / d_max if d_max > 0 else density
+                d_max = india_density.max()
+                india_density = india_density / d_max if d_max > 0 else india_density
         else:
-            d_max = density.max()
+            d_max = india_density.max()
             if d_max > 0:
-                density = density / d_max
+                india_density = india_density / d_max
 
+        density[india_mask] = india_density
         return density
 
     # ── POI Density ─────────────────────────────────────────────────────────