diff --git a/analyze.py b/analyze.py
new file mode 100644
index 0000000..f62c35b
--- /dev/null
+++ b/analyze.py
@@ -0,0 +1,597 @@
+import cv2
+import numpy as np
+
+
+class CompositionAnalyzer:
+    def __init__(self):
+        self.previous_gray = None
+        self.previous_saliency = None
+
+        self.cached_result = {
+            "score": 1.0,
+            "state": "green",
+            "motion": (0.0, 0.0),
+            "subject_center": (0.5, 0.5),
+            "direction": None,
+            "metrics": {},
+        }
+
+        self.motion_ema = np.array(
+            [0.0, 0.0],
+            dtype=np.float32,
+        )
+
+        self.centroid_ema = np.array(
+            [0.5, 0.5],
+            dtype=np.float32,
+        )
+
+        self.last_histogram = None
+
+    def analyze(
+        self,
+        frame,
+    ):
+        gray = cv2.cvtColor(
+            frame,
+            cv2.COLOR_BGR2GRAY,
+        )
+
+        frame_delta = self.compute_frame_delta(gray)
+
+        histogram_delta = self.compute_histogram_delta(gray)
+
+        if frame_delta < 1.25 and histogram_delta < 0.015:
+            return self.cached_result
+
+        saliency = self.compute_saliency(gray)
+
+        centroid = np.array(
+            self.compute_subject_centroid(saliency),
+            dtype=np.float32,
+        )
+
+        self.centroid_ema = self.centroid_ema * 0.84 + centroid * 0.16
+
+        motion_vector = np.array(
+            self.compute_optical_flow(gray),
+            dtype=np.float32,
+        )
+
+        self.motion_ema = self.motion_ema * 0.82 + motion_vector * 0.18
+
+        thirds_score = self.compute_thirds_score(self.centroid_ema)
+
+        edge_score = self.compute_edge_tension(self.centroid_ema)
+
+        balance_score = self.compute_balance(saliency)
+
+        negative_space = self.compute_negative_space(saliency)
+
+        clutter_score = self.compute_clutter(gray)
+
+        margin_score = self.compute_safe_margin(self.centroid_ema)
+
+        motion_score = self.compute_motion_stability(self.motion_ema)
+
+        drift_score = self.compute_predictive_drift(
+            self.centroid_ema,
+            self.motion_ema,
+        )
+
+        edge_density = self.compute_edge_density_balance(gray)
+
+        subject_isolation = self.compute_subject_isolation(saliency)
+
+        score = (
+            thirds_score * 0.18
+            + edge_score * 0.12
+            + balance_score * 0.11
+            + negative_space * 0.10
+            + clutter_score * 0.10
+            + margin_score * 0.12
+            + drift_score * 0.10
+            + motion_score * 0.07
+            + edge_density * 0.05
+            + subject_isolation * 0.05
+        )
+
+        score = float(
+            np.clip(
+                score,
+                0.0,
+                1.0,
+            )
+        )
+
+        if score >= 0.72:
+            state = "green"
+
+        elif score >= 0.48:
+            state = "yellow"
+
+        else:
+            state = "red"
+
+        direction = self.compute_directional_feedback(
+            self.centroid_ema,
+            self.motion_ema,
+        )
+
+        result = {
+            "score": score,
+            "state": state,
+            "motion": (
+                float(self.motion_ema[0]),
+                float(self.motion_ema[1]),
+            ),
+            "subject_center": (
+                float(self.centroid_ema[0]),
+                float(self.centroid_ema[1]),
+            ),
+            "direction": direction,
+            "metrics": {
+                "thirds": thirds_score,
+                "edge": edge_score,
+                "balance": balance_score,
+                "negative_space": negative_space,
+                "clutter": clutter_score,
+                "margin": margin_score,
+                "drift": drift_score,
+                "motion": motion_score,
+                "edge_density": edge_density,
+                "subject_isolation": subject_isolation,
+            },
+        }
+
+        self.previous_gray = gray
+        self.previous_saliency = saliency
+
+        self.cached_result = result
+
+        return result
+
+    def compute_frame_delta(
+        self,
+        gray,
+    ):
+        if self.previous_gray is None:
+            return 255.0
+
+        delta = cv2.absdiff(
+            self.previous_gray,
+            gray,
+        )
+
+        return float(np.mean(delta))
+
+    def compute_histogram_delta(
+        self,
+        gray,
+    ):
+        histogram = cv2.calcHist(
+            [gray],
+            [0],
+            None,
+            [64],
+            [0, 256],
+        )
+
+        histogram = cv2.normalize(
+            histogram,
+            histogram,
+        ).flatten()
+
+        if self.last_histogram is None:
+            self.last_histogram = histogram
+            return 1.0
+
+        delta = cv2.compareHist(
+            self.last_histogram.astype(np.float32),
+            histogram.astype(np.float32),
+            cv2.HISTCMP_BHATTACHARYYA,
+        )
+
+        self.last_histogram = histogram
+
+        return float(delta)
+
+    def compute_saliency(
+        self,
+        gray,
+    ):
+        blur = cv2.GaussianBlur(
+            gray,
+            (3, 3),
+            0,
+        )
+
+        spectrum = np.fft.fft2(blur)
+
+        log_amplitude = np.log(np.abs(spectrum) + 1e-6)
+
+        phase = np.angle(spectrum)
+
+        avg = cv2.blur(
+            log_amplitude,
+            (3, 3),
+        )
+
+        residual = log_amplitude - avg
+
+        spectral = np.exp(residual + 1j * phase)
+
+        saliency = np.abs(np.fft.ifft2(spectral)) ** 2
+
+        saliency = cv2.GaussianBlur(
+            saliency,
+            (9, 9),
+            2.5,
+        )
+
+        saliency = cv2.normalize(
+            saliency,
+            None,
+            0,
+            255,
+            cv2.NORM_MINMAX,
+        )
+
+        return saliency.astype(np.uint8)
+
+    def compute_subject_centroid(
+        self,
+        saliency,
+    ):
+        moments = cv2.moments(saliency)
+
+        if moments["m00"] <= 1:
+            return (
+                0.5,
+                0.5,
+            )
+
+        cx = moments["m10"] / moments["m00"]
+
+        cy = moments["m01"] / moments["m00"]
+
+        h, w = saliency.shape
+
+        return (
+            cx / w,
+            cy / h,
+        )
+
+    def compute_thirds_score(
+        self,
+        centroid,
+    ):
+        x, y = centroid
+
+        targets = (
+            (1 / 3, 1 / 3),
+            (2 / 3, 1 / 3),
+            (1 / 3, 2 / 3),
+            (2 / 3, 2 / 3),
+        )
+
+        distance = min(
+            np.hypot(
+                x - tx,
+                y - ty,
+            )
+            for tx, ty in targets
+        )
+
+        return float(
+            max(
+                0.0,
+                1.0 - distance * 2.4,
+            )
+        )
+
+    def compute_edge_tension(
+        self,
+        centroid,
+    ):
+        x, y = centroid
+
+        edge_distance = min(
+            x,
+            y,
+            1 - x,
+            1 - y,
+        )
+
+        return float(
+            min(
+                1.0,
+                edge_distance * 4,
+            )
+        )
+
+    # v4.5
+    def compute_balance(
+        self,
+        saliency,
+    ):
+        h, w = saliency.shape
+
+        left = np.sum(
+            saliency[:, : w // 2],
+            dtype=np.float64,
+        )
+
+        right = np.sum(
+            saliency[:, w // 2 :],
+            dtype=np.float64,
+        )
+
+        diff = abs(float(left) - float(right))
+
+        total = max(
+            1.0,
+            float(left + right),
+        )
+
+        return float(
+            np.clip(
+                1.0 - (diff / total),
+                0.0,
+                1.0,
+            )
+        )
+
+    def compute_negative_space(
+        self,
+        saliency,
+    ):
+        threshold = saliency > 32
+
+        active = np.count_nonzero(threshold)
+
+        total = saliency.size
+
+        occupied = active / total
+
+        return float(
+            max(
+                0.0,
+                1.0 - abs(occupied - 0.38),
+            )
+        )
+
+    def compute_clutter(
+        self,
+        gray,
+    ):
+        edges = cv2.Canny(
+            gray,
+            80,
+            160,
+        )
+
+        density = np.count_nonzero(edges) / edges.size
+
+        return float(
+            max(
+                0.0,
+                1.0 - density * 4,
+            )
+        )
+
+    def compute_safe_margin(
+        self,
+        centroid,
+    ):
+        x, y = centroid
+
+        margin = 0.08
+
+        if x < margin or y < margin or x > 1 - margin or y > 1 - margin:
+            return 0.1
+
+        return 1.0
+
+    def compute_optical_flow(
+        self,
+        gray,
+    ):
+        if self.previous_gray is None:
+            return (
+                0.0,
+                0.0,
+            )
+
+        points = cv2.goodFeaturesToTrack(
+            self.previous_gray,
+            maxCorners=64,
+            qualityLevel=0.01,
+            minDistance=12,
+        )
+
+        if points is None:
+            return (
+                0.0,
+                0.0,
+            )
+
+        next_points, status, _ = cv2.calcOpticalFlowPyrLK(
+            self.previous_gray,
+            gray,
+            points,
+            None,
+        )
+
+        if next_points is None or status is None:
+            return (
+                0.0,
+                0.0,
+            )
+
+        status = status.reshape(-1)
+
+        valid_old = points[status == 1]
+
+        valid_new = next_points[status == 1]
+
+        if len(valid_old) == 0 or len(valid_new) == 0:
+            return (
+                0.0,
+                0.0,
+            )
+
+        motion = (valid_new - valid_old).reshape(-1, 2)
+
+        if motion.size == 0:
+            return (
+                0.0,
+                0.0,
+            )
+
+        vector = np.mean(
+            motion,
+            axis=0,
+            dtype=np.float32,
+        )
+
+        return (
+            float(vector[0]),
+            float(vector[1]),
+        )
+
+    def compute_predictive_drift(
+        self,
+        centroid,
+        motion,
+    ):
+        x, y = centroid
+
+        mx, my = motion
+
+        future_x = x + (mx * 0.0008)
+
+        future_y = y + (my * 0.0008)
+
+        edge_distance = min(
+            future_x,
+            future_y,
+            1 - future_x,
+            1 - future_y,
+        )
+
+        return float(
+            max(
+                0.0,
+                min(
+                    1.0,
+                    edge_distance * 4,
+                ),
+            )
+        )
+
+    def compute_motion_stability(
+        self,
+        motion,
+    ):
+        magnitude = np.linalg.norm(motion)
+
+        return float(
+            max(
+                0.0,
+                1.0 - (magnitude / 42.0),
+            )
+        )
+
+    def compute_edge_density_balance(
+        self,
+        gray,
+    ):
+        edges = cv2.Canny(
+            gray,
+            100,
+            180,
+        )
+
+        h, w = edges.shape
+
+        quadrants = [
+            edges[: h // 2, : w // 2],
+            edges[: h // 2, w // 2 :],
+            edges[h // 2 :, : w // 2],
+            edges[h // 2 :, w // 2 :],
+        ]
+
+        densities = [np.count_nonzero(q) / q.size for q in quadrants]
+
+        std = np.std(densities)
+
+        return float(
+            max(
+                0.0,
+                1.0 - (std * 12.0),
+            )
+        )
+
+    def compute_subject_isolation(
+        self,
+        saliency,
+    ):
+        blurred = cv2.GaussianBlur(
+            saliency,
+            (31, 31),
+            0,
+        )
+
+        center_energy = np.max(blurred)
+
+        surrounding = np.mean(blurred)
+
+        if surrounding <= 0:
+            return 1.0
+
+        ratio = center_energy / surrounding
+
+        return float(
+            np.clip(
+                ratio / 4.0,
+                0.0,
+                1.0,
+            )
+        )
+
+    def compute_directional_feedback(
+        self,
+        centroid,
+        motion,
+    ):
+        x = float(centroid[0])
+
+        y = float(centroid[1])
+
+        mx = float(motion[0])
+
+        my = float(motion[1])
+
+        if x < 0.28:
+            return "move_left"
+
+        if x > 0.72:
+            return "move_right"
+
+        if y < 0.20:
+            return "reduce_headroom"
+
+        if y > 0.80:
+            return "recenter_subject"
+
+        if mx > 14:
+            return "stabilize_right_motion"
+
+        if mx < -14:
+            return "stabilize_left_motion"
+
+        if my > 14:
+            return "reduce_vertical_motion"
+
+        return None