diff --git a/.python-version b/.python-version
new file mode 100644
index 0000000..24ee5b1
--- /dev/null
+++ b/.python-version
@@ -0,0 +1 @@
+3.13
diff --git a/connect_four.js b/connect_four.js
index e2c4a32..4be09ac 100644
--- a/connect_four.js
+++ b/connect_four.js
@@ -166,12 +166,57 @@ function scanBoard(b) {
     return [0, []];
 }
 
+function evaluateBoard(b, aiP, huP) {
+    let score = 0;
+
+    // Center column bonus
+    for (let r = 0; r < ROWS; r++) {
+        if (b[3][r] === aiP) score += 3;
+        else if (b[3][r] === huP) score -= 3;
+    }
+
+    // Score a window of 4 cells by piece counts
+    function scoreWindow(c, r, dc, dr) {
+        let ai = 0, hu = 0;
+        for (let i = 0; i < 4; i++) {
+            const v = b[c + i * dc][r + i * dr];
+            if (v === aiP) ai++;
+            else if (v === huP) hu++;
+        }
+        if (ai > 0 && hu > 0) return 0;
+        if (ai === 3) return 50;
+        if (ai === 2) return 5;
+        if (hu === 3) return -50;
+        if (hu === 2) return -5;
+        return 0;
+    }
+
+    // Horizontal
+    for (let r = 0; r < ROWS; r++)
+        for (let c = 0; c <= COLS - 4; c++)
+            score += scoreWindow(c, r, 1, 0);
+    // Vertical
+    for (let r = 0; r <= ROWS - 4; r++)
+        for (let c = 0; c < COLS; c++)
+            score += scoreWindow(c, r, 0, 1);
+    // Diagonal up-right
+    for (let r = 0; r <= ROWS - 4; r++)
+        for (let c = 0; c <= COLS - 4; c++)
+            score += scoreWindow(c, r, 1, 1);
+    // Diagonal down-right
+    for (let r = 3; r < ROWS; r++)
+        for (let c = 0; c <= COLS - 4; c++)
+            score += scoreWindow(c, r, 1, -1);
+
+    return score;
+}
+
 // --- AI -----------------------------------------------------
 function minimax(b, depth, alpha, beta, isMax, aiP, huP) {
     const [winner] = scanBoard(b);
     if (winner === aiP) return 1000 + depth;
     if (winner === huP) return -1000 - depth;
-    if (depth === 0 || isBoardFull(b)) return 0;
+    if (depth === 0 || isBoardFull(b)) return evaluateBoard(b, aiP, huP);
 
     let best = isMax ? -10000 : 10000;
     for (const c of COL_ORDER) {
@@ -196,12 +241,19 @@ function performAiMove(b, aiP, lookAhead, isDemo = false, dPly = 4) {
     const huP = aiP === 1 ? 2 : 1;
     const ply = isDemo ? dPly : lookAhead;
 
-    // Phase 1: instant win / block
+    // Phase 1a: check ALL columns for instant AI win
     for (let c = 0; c < COLS; c++) {
         const r = getFirstEmptyRow(b, c);
         if (r === -1) continue;
         b[c][r] = aiP;
         if (scanBoard(b)[0] === aiP) { b[c][r] = 0; return c; }
+        b[c][r] = 0;
+    }
+
+    // Phase 1b: check ALL columns for opponent block
+    for (let c = 0; c < COLS; c++) {
+        const r = getFirstEmptyRow(b, c);
+        if (r === -1) continue;
         b[c][r] = huP;
         if (scanBoard(b)[0] === huP) { b[c][r] = 0; return c; }
         b[c][r] = 0;
diff --git a/pyproject.toml b/pyproject.toml
index 5f10da7..a8b1e27 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -2,11 +2,14 @@
 name = "connect-four-terminal"
 version = "1.0.0"
 description = "Connect Four terminal game with AI"
-requires-python = ">=3.10"
+requires-python = ">=3.10,<3.14"
 dependencies = [
     "rich>=13.0",
     "python-dotenv>=1.0",
     "readchar>=4.0",
+    "tensorflow>=2.16",
+    "numpy>=2.0",
+    "pygame>=2.5",
 ]
 
 [project.scripts]
diff --git a/rl/.python-version b/rl/.python-version
new file mode 100644
index 0000000..24ee5b1
--- /dev/null
+++ b/rl/.python-version
@@ -0,0 +1 @@
+3.13
diff --git a/rl/README.md b/rl/README.md
new file mode 100644
index 0000000..e69de29
diff --git a/rl/__init__.py b/rl/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/rl/__main__.py b/rl/__main__.py
new file mode 100644
index 0000000..1c84cac
--- /dev/null
+++ b/rl/__main__.py
@@ -0,0 +1,38 @@
+"""Entry point: python -m rl [train|export|info]"""
+
+import os
+import sys
+
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+
+def main():
+    cmd = sys.argv[1] if len(sys.argv) > 1 else "train"
+
+    if cmd == "train":
+        from .train import train
+        train()
+
+    elif cmd == "export":
+        from .export import export_tflite
+        model_path = sys.argv[2] if len(sys.argv) > 2 else "rl/checkpoints/model_final.keras"
+        export_tflite(model_path)
+
+    elif cmd == "visualize":
+        from .visualize import run_visualized
+        run_visualized()
+
+    elif cmd == "info":
+        from .model import build_model, print_model_info
+        model = build_model()
+        print_model_info(model)
+
+    else:
+        print(f"Unknown command: {cmd}")
+        print("Usage: python -m rl [train|visualize|export|info]")
+        sys.exit(1)
+
+
+main()
diff --git a/rl/config.py b/rl/config.py
new file mode 100644
index 0000000..786f8d2
--- /dev/null
+++ b/rl/config.py
@@ -0,0 +1,36 @@
+"""Training hyperparameters — edit these to tune your model."""
+
+# ── Model architecture ──────────────────────────────────────────────
+CONV_FILTERS = 32          # filters per conv layer (keep small for ESP32)
+NUM_CONV_LAYERS = 3        # number of convolutional blocks
+DENSE_UNITS = 64           # units in the dense layer before heads
+
+# ── Training ────────────────────────────────────────────────────────
+LEARNING_RATE = 1e-3       # Adam learning rate
+BATCH_SIZE = 256           # training batch size
+EPOCHS_PER_ITERATION = 4   # epochs per training iteration
+REPLAY_BUFFER_SIZE = 50000 # max samples kept in replay buffer
+
+# ── Self-play ───────────────────────────────────────────────────────
+NUM_ITERATIONS = 50        # total train iterations (self-play → train cycles)
+GAMES_PER_ITERATION = 100  # self-play games generated per iteration
+MCTS_SIMULATIONS = 50      # MCTS simulations per move
+MCTS_C_PUCT = 1.4          # exploration constant
+MCTS_TEMPERATURE = 1.0     # move selection temperature (1 = proportional, →0 = greedy)
+TEMP_DROP_MOVE = 10        # switch to greedy after this many moves
+
+# ── Parallelism ────────────────────────────────────────────────────
+NUM_WORKERS = 0                # 0 = use all available CPU cores
+
+# ── Reward shaping ──────────────────────────────────────────────────
+WIN_REWARD = 1.0
+DRAW_REWARD = 0.0
+LOSS_REWARD = -1.0
+
+# ── Checkpointing ──────────────────────────────────────────────────
+CHECKPOINT_DIR = "rl/checkpoints"
+CHECKPOINT_INTERVAL = 5    # save model every N iterations
+EXPORT_DIR = "rl/export"
+
+# ── ESP32 export ────────────────────────────────────────────────────
+QUANTIZE_INT8 = True       # int8 quantization for TFLite (recommended for ESP32)
diff --git a/rl/export.py b/rl/export.py
new file mode 100644
index 0000000..be15365
--- /dev/null
+++ b/rl/export.py
@@ -0,0 +1,86 @@
+"""Export trained Keras model to TFLite (optionally int8-quantized) for ESP32."""
+
+import os
+import numpy as np
+
+from .game import ConnectFour, ROWS, COLS
+from .config import EXPORT_DIR, QUANTIZE_INT8
+
+
+def representative_dataset():
+    """Generate sample inputs for int8 calibration."""
+    game = ConnectFour()
+    for _ in range(200):
+        game.reset()
+        # Play random moves to get diverse board states
+        moves = np.random.randint(0, min(ROWS * COLS, 20))
+        for _ in range(moves):
+            legal = game.legal_moves()
+            if not legal or game.done:
+                break
+            game.step(np.random.choice(legal))
+        yield [game.get_state()[np.newaxis].astype(np.float32)]
+
+
+def export_tflite(model_path, quantize=None):
+    """Convert a saved Keras model to TFLite.
+
+    Args:
+        model_path: Path to the .keras model file.
+        quantize: Override quantization setting. If None, uses config.QUANTIZE_INT8.
+    """
+    import tensorflow as tf
+
+    if quantize is None:
+        quantize = QUANTIZE_INT8
+
+    os.makedirs(EXPORT_DIR, exist_ok=True)
+
+    model = tf.keras.models.load_model(model_path)
+
+    converter = tf.lite.TFLiteConverter.from_keras_model(model)
+
+    if quantize:
+        converter.optimizations = [tf.lite.Optimize.DEFAULT]
+        converter.representative_dataset = representative_dataset
+        converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+        converter.inference_input_type = tf.int8
+        converter.inference_output_type = tf.int8
+        suffix = "_int8"
+    else:
+        suffix = "_f32"
+
+    tflite_model = converter.convert()
+
+    out_path = os.path.join(EXPORT_DIR, f"connect4{suffix}.tflite")
+    with open(out_path, "wb") as f:
+        f.write(tflite_model)
+
+    size_kb = len(tflite_model) / 1024
+    print(f"Exported: {out_path} ({size_kb:.1f} KB)")
+
+    # Also export as C header for direct embedding in firmware
+    header_path = os.path.join(EXPORT_DIR, f"connect4_model{suffix}.h")
+    _write_c_header(tflite_model, header_path)
+    print(f"C header: {header_path}")
+
+    return out_path
+
+
+def _write_c_header(model_bytes, path):
+    """Write TFLite model as a C byte array for ESP32 firmware inclusion."""
+    with open(path, "w") as f:
+        f.write("#pragma once\n\n")
+        f.write(f"// Auto-generated — {len(model_bytes)} bytes\n")
+        f.write(f"const unsigned int connect4_model_len = {len(model_bytes)};\n")
+        f.write("alignas(16) const unsigned char connect4_model[] = {\n")
+        for i in range(0, len(model_bytes), 12):
+            chunk = model_bytes[i:i + 12]
+            f.write("    " + ", ".join(f"0x{b:02x}" for b in chunk) + ",\n")
+        f.write("};\n")
+
+
+if __name__ == "__main__":
+    import sys
+    model_path = sys.argv[1] if len(sys.argv) > 1 else "rl/checkpoints/model_final.keras"
+    export_tflite(model_path)
diff --git a/rl/game.py b/rl/game.py
new file mode 100644
index 0000000..c95153f
--- /dev/null
+++ b/rl/game.py
@@ -0,0 +1,102 @@
+"""Connect Four game environment for self-play training."""
+
+import numpy as np
+
+ROWS = 6
+COLS = 7
+WIN_LENGTH = 4
+
+
+class ConnectFour:
+    """Connect Four game with numpy board representation.
+
+    Board encoding: 0 = empty, 1 = player 1, -1 = player 2.
+    """
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.board = np.zeros((ROWS, COLS), dtype=np.int8)
+        self.current_player = 1
+        self.done = False
+        self.winner = 0  # 0 = no winner / draw, 1 or -1
+        self.move_count = 0
+        return self.get_state()
+
+    def get_state(self):
+        """Return board from current player's perspective as (6,7,2) tensor.
+
+        Channel 0: current player's pieces (1s).
+        Channel 1: opponent's pieces (1s).
+        """
+        state = np.zeros((ROWS, COLS, 2), dtype=np.float32)
+        state[:, :, 0] = (self.board == self.current_player).astype(np.float32)
+        state[:, :, 1] = (self.board == -self.current_player).astype(np.float32)
+        return state
+
+    def legal_moves(self):
+        """Return list of columns that are not full."""
+        return [c for c in range(COLS) if self.board[0, c] == 0]
+
+    def legal_moves_mask(self):
+        """Return binary mask of legal columns."""
+        return (self.board[0] == 0).astype(np.float32)
+
+    def step(self, col):
+        """Play a move in the given column. Returns (state, reward, done)."""
+        if self.done:
+            raise ValueError("Game is already over.")
+        if col < 0 or col >= COLS or self.board[0, col] != 0:
+            raise ValueError(f"Illegal move: column {col}")
+
+        # Drop piece
+        row = self._get_drop_row(col)
+        self.board[row, col] = self.current_player
+        self.move_count += 1
+
+        # Check win
+        if self._check_win(row, col):
+            self.done = True
+            self.winner = self.current_player
+            reward = 1.0
+        elif self.move_count == ROWS * COLS:
+            self.done = True
+            self.winner = 0
+            reward = 0.0
+        else:
+            reward = 0.0
+
+        # Switch player
+        self.current_player *= -1
+        return self.get_state(), reward, self.done
+
+    def _get_drop_row(self, col):
+        for r in range(ROWS - 1, -1, -1):
+            if self.board[r, col] == 0:
+                return r
+        raise ValueError(f"Column {col} is full")
+
+    def _check_win(self, row, col):
+        player = self.board[row, col]
+        directions = [(0, 1), (1, 0), (1, 1), (1, -1)]
+        for dr, dc in directions:
+            count = 1
+            for sign in (1, -1):
+                r, c = row + sign * dr, col + sign * dc
+                while 0 <= r < ROWS and 0 <= c < COLS and self.board[r, c] == player:
+                    count += 1
+                    r += sign * dr
+                    c += sign * dc
+            if count >= WIN_LENGTH:
+                return True
+        return False
+
+    def clone(self):
+        g = ConnectFour()
+        g.board = self.board.copy()
+        g.current_player = self.current_player
+        g.done = self.done
+        g.winner = self.winner
+        g.move_count = self.move_count
+        return g
diff --git a/rl/main.py b/rl/main.py
new file mode 100644
index 0000000..f118498
--- /dev/null
+++ b/rl/main.py
@@ -0,0 +1,6 @@
+def main():
+    print("Hello from rl!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/rl/mcts.py b/rl/mcts.py
new file mode 100644
index 0000000..2b56835
--- /dev/null
+++ b/rl/mcts.py
@@ -0,0 +1,103 @@
+"""Monte Carlo Tree Search for self-play data generation."""
+
+import math
+import numpy as np
+from .game import ConnectFour
+from .config import MCTS_C_PUCT
+
+
+class MCTSNode:
+    __slots__ = ("parent", "action", "prior", "visit_count", "value_sum", "children", "game")
+
+    def __init__(self, game, parent=None, action=None, prior=0.0):
+        self.game = game
+        self.parent = parent
+        self.action = action
+        self.prior = prior
+        self.visit_count = 0
+        self.value_sum = 0.0
+        self.children = {}
+
+    @property
+    def q_value(self):
+        if self.visit_count == 0:
+            return 0.0
+        return self.value_sum / self.visit_count
+
+    def ucb_score(self):
+        parent_visits = self.parent.visit_count if self.parent else 1
+        exploration = MCTS_C_PUCT * self.prior * math.sqrt(parent_visits) / (1 + self.visit_count)
+        return self.q_value + exploration
+
+    def is_leaf(self):
+        return len(self.children) == 0
+
+    def expand(self, policy_probs):
+        """Expand node using network policy output."""
+        legal = self.game.legal_moves()
+        for col in legal:
+            if col not in self.children:
+                self.children[col] = MCTSNode(
+                    game=None, parent=self, action=col, prior=policy_probs[col]
+                )
+
+    def select_child(self):
+        return max(self.children.values(), key=lambda c: c.ucb_score())
+
+
+def run_mcts(game, model, num_simulations):
+    """Run MCTS from current game state, return visit-count policy vector."""
+    root = MCTSNode(game.clone())
+
+    # Evaluate root
+    state = root.game.get_state()
+    policy_logits, value = model.predict(state[np.newaxis], verbose=0)
+    policy = _mask_and_normalize(policy_logits[0], root.game.legal_moves_mask())
+    root.expand(policy)
+
+    for _ in range(num_simulations):
+        node = root
+        sim_game = root.game.clone()
+
+        # SELECT — walk down tree picking best UCB child
+        while not node.is_leaf() and not sim_game.done:
+            node = node.select_child()
+            sim_game.step(node.action)
+
+        # EVALUATE leaf
+        if sim_game.done:
+            # Terminal: value from perspective of player who just moved
+            if sim_game.winner == 0:
+                leaf_value = 0.0
+            else:
+                # The winner is sim_game.winner; current_player already switched
+                leaf_value = -1.0  # current player lost (winner was previous player)
+        else:
+            node.game = sim_game.clone()
+            state = sim_game.get_state()
+            policy_logits, value = model.predict(state[np.newaxis], verbose=0)
+            leaf_value = value[0, 0]
+            policy = _mask_and_normalize(policy_logits[0], sim_game.legal_moves_mask())
+            node.expand(policy)
+
+        # BACKUP — propagate value up, flipping sign each level
+        while node is not None:
+            node.visit_count += 1
+            node.value_sum += leaf_value
+            leaf_value = -leaf_value
+            node = node.parent
+
+    # Build policy from visit counts
+    visits = np.zeros(7, dtype=np.float32)
+    for col, child in root.children.items():
+        visits[col] = child.visit_count
+    return visits
+
+
+def _mask_and_normalize(logits, mask):
+    """Apply legal-move mask and softmax."""
+    logits = np.array(logits, dtype=np.float64)
+    logits[mask == 0] = -1e9
+    exp = np.exp(logits - np.max(logits))
+    probs = exp / np.sum(exp)
+    return probs.astype(np.float32)
diff --git a/rl/model.py b/rl/model.py
new file mode 100644
index 0000000..ed8acc8
--- /dev/null
+++ b/rl/model.py
@@ -0,0 +1,54 @@
+"""Compact dual-head neural network (policy + value) sized for ESP32."""
+
+from .config import CONV_FILTERS, NUM_CONV_LAYERS, DENSE_UNITS, LEARNING_RATE
+
+
+def build_model():
+    """Build a small AlphaZero-style network.
+
+    Input:  (6, 7, 2)  — current player pieces / opponent pieces
+    Output: policy (7,) — log-probabilities over columns
+            value  (1,) — board evaluation in [-1, 1]
+    """
+    from tensorflow import keras
+    from tensorflow.keras import layers
+
+    inp = layers.Input(shape=(6, 7, 2), name="board")
+
+    x = inp
+    for i in range(NUM_CONV_LAYERS):
+        x = layers.Conv2D(
+            CONV_FILTERS, 3, padding="same", activation="relu", name=f"conv{i}"
+        )(x)
+        x = layers.BatchNormalization(name=f"bn{i}")(x)
+
+    flat = layers.Flatten(name="flat")(x)
+    shared = layers.Dense(DENSE_UNITS, activation="relu", name="shared_dense")(flat)
+
+    # Policy head
+    policy = layers.Dense(7, name="policy_logits")(shared)
+
+    # Value head
+    value = layers.Dense(1, activation="tanh", name="value")(shared)
+
+    model = keras.Model(inputs=inp, outputs=[policy, value], name="connect4_net")
+
+    model.compile(
+        optimizer=keras.optimizers.Adam(learning_rate=LEARNING_RATE),
+        loss={
+            "policy_logits": keras.losses.CategoricalCrossentropy(from_logits=True),
+            "value": keras.losses.MeanSquaredError(),
+        },
+        loss_weights={"policy_logits": 1.0, "value": 1.0},
+    )
+    return model
+
+
+def print_model_info(model):
+    model.summary()
+    total_params = model.count_params()
+    approx_size_kb = total_params * 4 / 1024  # float32
+    approx_int8_kb = total_params / 1024       # int8
+    print(f"\nTotal parameters: {total_params:,}")
+    print(f"Approx size (float32): {approx_size_kb:.1f} KB")
+    print(f"Approx size (int8):    {approx_int8_kb:.1f} KB")
diff --git a/rl/train.py b/rl/train.py
new file mode 100644
index 0000000..8d10045
--- /dev/null
+++ b/rl/train.py
@@ -0,0 +1,143 @@
+"""Self-play training loop with parallel game generation."""
+
+import os
+import numpy as np
+from collections import deque
+from multiprocessing import Pool, cpu_count
+
+from .game import ConnectFour
+from .model import build_model, print_model_info
+from .mcts import run_mcts
+from .config import (
+    NUM_ITERATIONS, GAMES_PER_ITERATION, MCTS_SIMULATIONS,
+    MCTS_TEMPERATURE, TEMP_DROP_MOVE,
+    WIN_REWARD, DRAW_REWARD, LOSS_REWARD,
+    BATCH_SIZE, EPOCHS_PER_ITERATION, REPLAY_BUFFER_SIZE,
+    CHECKPOINT_DIR, CHECKPOINT_INTERVAL, NUM_WORKERS,
+)
+
+# Per-worker global model (loaded once per process)
+_worker_model = None
+
+
+def _init_worker(weights_list):
+    """Initialize a worker process with its own model copy."""
+    global _worker_model
+    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+    os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"
+    os.environ["CUDA_VISIBLE_DEVICES"] = ""
+    _worker_model = build_model()
+    _worker_model.set_weights(weights_list)
+
+
+def _play_one_game(_):
+    """Play a single self-play game in a worker process."""
+    game = ConnectFour()
+    trajectory = []
+
+    while not game.done:
+        state = game.get_state()
+        visit_counts = run_mcts(game, _worker_model, MCTS_SIMULATIONS)
+
+        if game.move_count < TEMP_DROP_MOVE:
+            temp = MCTS_TEMPERATURE
+        else:
+            temp = 0.1
+
+        if temp < 0.2:
+            action = int(np.argmax(visit_counts))
+            policy = np.zeros(7, dtype=np.float32)
+            policy[action] = 1.0
+        else:
+            counts = visit_counts ** (1.0 / temp)
+            policy = counts / counts.sum()
+            action = np.random.choice(7, p=policy)
+
+        trajectory.append((state, policy, game.current_player))
+        game.step(action)
+
+    samples = []
+    for state, policy, player in trajectory:
+        if game.winner == 0:
+            value = DRAW_REWARD
+        elif game.winner == player:
+            value = WIN_REWARD
+        else:
+            value = LOSS_REWARD
+        samples.append((state, policy, value))
+
+    return samples
+
+
+def train():
+    """Main training entry point."""
+    model = build_model()
+    print_model_info(model)
+
+    num_workers = NUM_WORKERS if NUM_WORKERS > 0 else cpu_count()
+    print(f"Using {num_workers} worker processes for self-play")
+
+    replay_buffer = deque(maxlen=REPLAY_BUFFER_SIZE)
+    os.makedirs(CHECKPOINT_DIR, exist_ok=True)
+
+    for iteration in range(1, NUM_ITERATIONS + 1):
+        print(f"\n{'='*60}")
+        print(f"Iteration {iteration}/{NUM_ITERATIONS}")
+        print(f"{'='*60}")
+
+        # ── Self-play (parallel) ───────────────────────────────
+        weights = model.get_weights()
+        with Pool(processes=num_workers, initializer=_init_worker, initargs=(weights,)) as pool:
+            results = pool.map(_play_one_game, range(GAMES_PER_ITERATION))
+
+        wins = {1: 0, -1: 0, 0: 0}
+        for samples in results:
+            replay_buffer.extend(samples)
+            if samples:
+                last_value = samples[-1][2]
+                if last_value == WIN_REWARD:
+                    wins[1] += 1
+                elif last_value == LOSS_REWARD:
+                    wins[-1] += 1
+                else:
+                    wins[0] += 1
+
+        print(f"  Self-play: {GAMES_PER_ITERATION} games "
+              f"(P1 wins: {wins[1]}, P2 wins: {wins[-1]}, draws: {wins[0]})")
+        print(f"  Buffer size: {len(replay_buffer)}")
+
+        # ── Train ───────────────────────────────────────────────
+        if len(replay_buffer) >= BATCH_SIZE:
+            sample_size = min(len(replay_buffer), BATCH_SIZE * EPOCHS_PER_ITERATION)
+            indices = np.random.choice(len(replay_buffer), size=sample_size, replace=False)
+            batch = [replay_buffer[i] for i in indices]
+
+            states = np.array([s[0] for s in batch])
+            policies = np.array([s[1] for s in batch])
+            values = np.array([s[2] for s in batch]).reshape(-1, 1)
+
+            history = model.fit(
+                states,
+                {"policy_logits": policies, "value": values},
+                batch_size=BATCH_SIZE,
+                epochs=EPOCHS_PER_ITERATION,
+                verbose=1,
+            )
+            policy_loss = history.history["policy_logits_loss"][-1]
+            value_loss = history.history["value_loss"][-1]
+            print(f"  Policy loss: {policy_loss:.4f}  Value loss: {value_loss:.4f}")
+
+        # ── Checkpoint ──────────────────────────────────────────
+        if iteration % CHECKPOINT_INTERVAL == 0:
+            path = os.path.join(CHECKPOINT_DIR, f"model_iter{iteration}.keras")
+            model.save(path)
+            print(f"  Saved checkpoint: {path}")
+
+    final_path = os.path.join(CHECKPOINT_DIR, "model_final.keras")
+    model.save(final_path)
+    print(f"\nTraining complete. Final model saved to {final_path}")
+    return model
+
+
+if __name__ == "__main__":
+    train()
diff --git a/rl/visualize.py b/rl/visualize.py
new file mode 100644
index 0000000..7ab6261
--- /dev/null
+++ b/rl/visualize.py
@@ -0,0 +1,482 @@
+"""Pygame visualization of Connect Four RL training.
+
+Left panel:  live self-play game board
+Right panel: loss curves + win-rate chart + training stats
+"""
+
+import os
+import threading
+import time
+from collections import deque
+
+import numpy as np
+import pygame
+
+from .game import ConnectFour, ROWS, COLS
+from .model import build_model, print_model_info
+from .mcts import run_mcts
+from .config import (
+    NUM_ITERATIONS, GAMES_PER_ITERATION, MCTS_SIMULATIONS,
+    MCTS_TEMPERATURE, TEMP_DROP_MOVE,
+    WIN_REWARD, DRAW_REWARD, LOSS_REWARD,
+    BATCH_SIZE, EPOCHS_PER_ITERATION, REPLAY_BUFFER_SIZE,
+    CHECKPOINT_DIR, CHECKPOINT_INTERVAL, NUM_WORKERS,
+)
+from multiprocessing import Pool, cpu_count
+
+# ── Layout constants ────────────────────────────────────────────────
+CELL = 80
+BOARD_W = COLS * CELL
+BOARD_H = ROWS * CELL
+PANEL_W = 420
+MARGIN = 20
+WIN_W = BOARD_W + PANEL_W + MARGIN * 3
+WIN_H = BOARD_H + MARGIN * 2
+FPS = 30
+
+# ── Colors ──────────────────────────────────────────────────────────
+BG = (30, 30, 40)
+BOARD_BG = (0, 60, 180)
+EMPTY = (20, 20, 30)
+P1_COLOR = (255, 220, 50)   # yellow
+P2_COLOR = (220, 40, 40)    # red
+WIN_HIGHLIGHT = (100, 255, 100)
+GRID_LINE = (0, 40, 140)
+TEXT_COLOR = (220, 220, 220)
+CHART_BG = (40, 40, 55)
+POLICY_LINE = (80, 200, 255)
+VALUE_LINE = (255, 160, 60)
+P1_CHART = (255, 220, 50)
+P2_CHART = (220, 40, 40)
+DRAW_CHART = (140, 140, 140)
+
+# ── Shared state between training thread and pygame loop ────────────
+_state = {
+    "board": np.zeros((ROWS, COLS), dtype=np.int8),
+    "iteration": 0,
+    "game_num": 0,
+    "phase": "init",         # init / self-play / training / done
+    "policy_losses": [],
+    "value_losses": [],
+    "win_history": [],       # list of (p1_wins, p2_wins, draws) per iteration
+    "move_delay": 0.3,
+    "status": "Initializing...",
+    "winner": 0,
+    "running": True,
+}
+_lock = threading.Lock()
+
+
+# ── Worker setup (same as train.py) ─────────────────────────────────
+_worker_model = None
+
+
+def _init_worker(weights_list):
+    global _worker_model
+    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+    os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"
+    os.environ["CUDA_VISIBLE_DEVICES"] = ""
+    _worker_model = build_model()
+    _worker_model.set_weights(weights_list)
+
+
+def _play_one_game(_):
+    game = ConnectFour()
+    trajectory = []
+    while not game.done:
+        state = game.get_state()
+        visit_counts = run_mcts(game, _worker_model, MCTS_SIMULATIONS)
+        if game.move_count < TEMP_DROP_MOVE:
+            temp = MCTS_TEMPERATURE
+        else:
+            temp = 0.1
+        if temp < 0.2:
+            action = int(np.argmax(visit_counts))
+            policy = np.zeros(7, dtype=np.float32)
+            policy[action] = 1.0
+        else:
+            counts = visit_counts ** (1.0 / temp)
+            policy = counts / counts.sum()
+            action = np.random.choice(7, p=policy)
+        trajectory.append((state, policy, game.current_player))
+        game.step(action)
+    samples = []
+    for state, policy, player in trajectory:
+        if game.winner == 0:
+            value = DRAW_REWARD
+        elif game.winner == player:
+            value = WIN_REWARD
+        else:
+            value = LOSS_REWARD
+        samples.append((state, policy, value))
+    return samples
+
+
+def _play_showcase_game(model):
+    """Play one game slowly on the main training thread, updating shared state."""
+    game = ConnectFour()
+    trajectory = []
+
+    with _lock:
+        _state["board"] = game.board.copy()
+        _state["winner"] = 0
+
+    while not game.done and _state["running"]:
+        state = game.get_state()
+        visit_counts = run_mcts(game, model, MCTS_SIMULATIONS)
+
+        if game.move_count < TEMP_DROP_MOVE:
+            temp = MCTS_TEMPERATURE
+        else:
+            temp = 0.1
+        if temp < 0.2:
+            action = int(np.argmax(visit_counts))
+            policy = np.zeros(7, dtype=np.float32)
+            policy[action] = 1.0
+        else:
+            counts = visit_counts ** (1.0 / temp)
+            policy = counts / counts.sum()
+            action = np.random.choice(7, p=policy)
+
+        trajectory.append((state, policy, game.current_player))
+        game.step(action)
+
+        with _lock:
+            _state["board"] = game.board.copy()
+
+        time.sleep(_state["move_delay"])
+
+    with _lock:
+        _state["winner"] = game.winner
+
+    samples = []
+    for state, policy, player in trajectory:
+        if game.winner == 0:
+            value = DRAW_REWARD
+        elif game.winner == player:
+            value = WIN_REWARD
+        else:
+            value = LOSS_REWARD
+        samples.append((state, policy, value))
+    return samples
+
+
+def _training_thread():
+    """Run the full training loop, pushing updates to shared state."""
+    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+    model = build_model()
+    print_model_info(model)
+
+    num_workers = NUM_WORKERS if NUM_WORKERS > 0 else cpu_count()
+    replay_buffer = deque(maxlen=REPLAY_BUFFER_SIZE)
+    os.makedirs(CHECKPOINT_DIR, exist_ok=True)
+
+    with _lock:
+        _state["status"] = f"Using {num_workers} workers"
+
+    for iteration in range(1, NUM_ITERATIONS + 1):
+        if not _state["running"]:
+            break
+
+        with _lock:
+            _state["iteration"] = iteration
+            _state["phase"] = "self-play"
+            _state["status"] = f"Iteration {iteration}/{NUM_ITERATIONS} - Self-play"
+
+        # Play one showcase game visually
+        with _lock:
+            _state["game_num"] = 0
+        showcase_samples = _play_showcase_game(model)
+        replay_buffer.extend(showcase_samples)
+
+        # Play remaining games in parallel
+        remaining = GAMES_PER_ITERATION - 1
+        if remaining > 0 and _state["running"]:
+            with _lock:
+                _state["status"] = f"Iter {iteration} - Playing {remaining} games (parallel)..."
+
+            weights = model.get_weights()
+            with Pool(processes=num_workers, initializer=_init_worker, initargs=(weights,)) as pool:
+                results = pool.map(_play_one_game, range(remaining))
+
+            for samples in results:
+                replay_buffer.extend(samples)
+
+        # Count wins across all games this iteration
+        wins = {1: 0, -1: 0, 0: 0}
+        # Showcase game
+        if showcase_samples:
+            last_val = showcase_samples[-1][2]
+            if last_val == WIN_REWARD:
+                wins[1] += 1
+            elif last_val == LOSS_REWARD:
+                wins[-1] += 1
+            else:
+                wins[0] += 1
+        # Parallel games
+        if remaining > 0 and _state["running"]:
+            for samples in results:
+                if samples:
+                    last_val = samples[-1][2]
+                    if last_val == WIN_REWARD:
+                        wins[1] += 1
+                    elif last_val == LOSS_REWARD:
+                        wins[-1] += 1
+                    else:
+                        wins[0] += 1
+
+        with _lock:
+            _state["win_history"].append((wins[1], wins[-1], wins[0]))
+
+        # Train
+        if len(replay_buffer) >= BATCH_SIZE and _state["running"]:
+            with _lock:
+                _state["phase"] = "training"
+                _state["status"] = f"Iter {iteration} - Training..."
+
+            sample_size = min(len(replay_buffer), BATCH_SIZE * EPOCHS_PER_ITERATION)
+            indices = np.random.choice(len(replay_buffer), size=sample_size, replace=False)
+            batch = [replay_buffer[i] for i in indices]
+
+            states = np.array([s[0] for s in batch])
+            policies = np.array([s[1] for s in batch])
+            values = np.array([s[2] for s in batch]).reshape(-1, 1)
+
+            history = model.fit(
+                states,
+                {"policy_logits": policies, "value": values},
+                batch_size=BATCH_SIZE,
+                epochs=EPOCHS_PER_ITERATION,
+                verbose=0,
+            )
+
+            with _lock:
+                _state["policy_losses"].append(history.history["policy_logits_loss"][-1])
+                _state["value_losses"].append(history.history["value_loss"][-1])
+
+        # Checkpoint
+        if iteration % CHECKPOINT_INTERVAL == 0:
+            path = os.path.join(CHECKPOINT_DIR, f"model_iter{iteration}.keras")
+            model.save(path)
+
+    if _state["running"]:
+        final_path = os.path.join(CHECKPOINT_DIR, "model_final.keras")
+        model.save(final_path)
+
+    with _lock:
+        _state["phase"] = "done"
+        _state["status"] = "Training complete!"
+
+
+# ── Drawing helpers ─────────────────────────────────────────────────
+
+def _draw_board(surface, board, x0, y0):
+    """Draw the Connect Four board."""
+    # Board background
+    pygame.draw.rect(surface, BOARD_BG, (x0, y0, BOARD_W, BOARD_H), border_radius=8)
+
+    for r in range(ROWS):
+        for c in range(COLS):
+            cx = x0 + c * CELL + CELL // 2
+            cy = y0 + r * CELL + CELL // 2
+            radius = CELL // 2 - 6
+
+            val = board[r, c]
+            if val == 1:
+                color = P1_COLOR
+            elif val == -1:
+                color = P2_COLOR
+            else:
+                color = EMPTY
+
+            pygame.draw.circle(surface, color, (cx, cy), radius)
+            pygame.draw.circle(surface, GRID_LINE, (cx, cy), radius, 2)
+
+
+def _draw_chart(surface, x, y, w, h, series_list, colors, title, font):
+    """Draw a simple line chart with multiple series."""
+    pygame.draw.rect(surface, CHART_BG, (x, y, w, h), border_radius=6)
+    pygame.draw.rect(surface, (60, 60, 75), (x, y, w, h), 1, border_radius=6)
+
+    # Title
+    title_surf = font.render(title, True, TEXT_COLOR)
+    surface.blit(title_surf, (x + 8, y + 4))
+
+    chart_x = x + 8
+    chart_y = y + 24
+    chart_w = w - 16
+    chart_h = h - 32
+
+    if not any(series_list):
+        return
+
+    # Find global min/max
+    all_vals = [v for s in series_list if s for v in s]
+    if not all_vals:
+        return
+    min_val = min(all_vals)
+    max_val = max(all_vals)
+    val_range = max_val - min_val if max_val != min_val else 1.0
+
+    for series, color in zip(series_list, colors):
+        if len(series) < 2:
+            continue
+        points = []
+        for i, v in enumerate(series):
+            px = chart_x + int(i / (len(series) - 1) * chart_w)
+            py = chart_y + chart_h - int((v - min_val) / val_range * chart_h)
+            points.append((px, py))
+        pygame.draw.lines(surface, color, False, points, 2)
+
+
+def _draw_stacked_bar(surface, x, y, w, h, win_history, font):
+    """Draw stacked bar chart of win rates."""
+    pygame.draw.rect(surface, CHART_BG, (x, y, w, h), border_radius=6)
+    pygame.draw.rect(surface, (60, 60, 75), (x, y, w, h), 1, border_radius=6)
+
+    title_surf = font.render("Win rates per iteration", True, TEXT_COLOR)
+    surface.blit(title_surf, (x + 8, y + 4))
+
+    if not win_history:
+        return
+
+    chart_x = x + 8
+    chart_y = y + 24
+    chart_w = w - 16
+    chart_h = h - 48
+
+    n = len(win_history)
+    bar_w = max(2, chart_w // max(n, 1))
+
+    for i, (p1, p2, dr) in enumerate(win_history):
+        total = p1 + p2 + dr
+        if total == 0:
+            continue
+        bx = chart_x + int(i / max(n, 1) * chart_w)
+
+        # Stack: P1 (bottom), draws (middle), P2 (top)
+        h1 = int(p1 / total * chart_h)
+        hd = int(dr / total * chart_h)
+        h2 = chart_h - h1 - hd
+
+        by = chart_y
+        pygame.draw.rect(surface, P2_CHART, (bx, by, bar_w - 1, h2))
+        by += h2
+        pygame.draw.rect(surface, DRAW_CHART, (bx, by, bar_w - 1, hd))
+        by += hd
+        pygame.draw.rect(surface, P1_CHART, (bx, by, bar_w - 1, h1))
+
+    # Legend
+    ly = y + h - 18
+    for label, color, lx in [("P1", P1_CHART, x + 8), ("Draw", DRAW_CHART, x + 70), ("P2", P2_CHART, x + 150)]:
+        pygame.draw.rect(surface, color, (lx, ly, 12, 12))
+        surface.blit(font.render(label, True, TEXT_COLOR), (lx + 16, ly - 2))
+
+
+def run_visualized():
+    """Launch pygame window and run training with live visualization."""
+    pygame.init()
+    screen = pygame.display.set_mode((WIN_W, WIN_H))
+    pygame.display.set_caption("Connect Four RL Training")
+    clock = pygame.time.Clock()
+    font = pygame.font.SysFont("monospace", 14)
+    font_big = pygame.font.SysFont("monospace", 18, bold=True)
+
+    # Start training in background thread
+    train_thread = threading.Thread(target=_training_thread, daemon=True)
+    train_thread.start()
+
+    running = True
+    while running:
+        for event in pygame.event.get():
+            if event.type == pygame.QUIT:
+                running = False
+                _state["running"] = False
+            elif event.type == pygame.KEYDOWN:
+                if event.key == pygame.K_ESCAPE:
+                    running = False
+                    _state["running"] = False
+                elif event.key == pygame.K_UP:
+                    _state["move_delay"] = max(0.05, _state["move_delay"] - 0.05)
+                elif event.key == pygame.K_DOWN:
+                    _state["move_delay"] = min(2.0, _state["move_delay"] + 0.05)
+
+        screen.fill(BG)
+
+        with _lock:
+            board = _state["board"].copy()
+            iteration = _state["iteration"]
+            phase = _state["phase"]
+            status = _state["status"]
+            policy_losses = list(_state["policy_losses"])
+            value_losses = list(_state["value_losses"])
+            win_history = list(_state["win_history"])
+            winner = _state["winner"]
+            delay = _state["move_delay"]
+
+        # ── Left: game board ────────────────────────────────────
+        bx, by = MARGIN, MARGIN
+        _draw_board(screen, board, bx, by)
+
+        # Winner overlay
+        if winner != 0 and phase == "self-play":
+            label = f"Player {1 if winner == 1 else 2} wins!"
+            color = P1_COLOR if winner == 1 else P2_COLOR
+            win_surf = font_big.render(label, True, color)
+            wrect = win_surf.get_rect(center=(bx + BOARD_W // 2, by + BOARD_H + 2))
+            if wrect.bottom < WIN_H:
+                screen.blit(win_surf, wrect)
+
+        # ── Right panel ────────────────────────────────────────
+        px = BOARD_W + MARGIN * 2
+        py = MARGIN
+
+        # Status
+        status_surf = font_big.render(status, True, TEXT_COLOR)
+        screen.blit(status_surf, (px, py))
+        py += 28
+
+        iter_surf = font.render(f"Iteration: {iteration}/{NUM_ITERATIONS}  Phase: {phase}", True, TEXT_COLOR)
+        screen.blit(iter_surf, (px, py))
+        py += 20
+
+        delay_surf = font.render(f"Move delay: {delay:.2f}s  (Up/Down to adjust)", True, (150, 150, 170))
+        screen.blit(delay_surf, (px, py))
+        py += 28
+
+        # Loss chart
+        chart_h = 140
+        _draw_chart(
+            screen, px, py, PANEL_W, chart_h,
+            [policy_losses, value_losses],
+            [POLICY_LINE, VALUE_LINE],
+            "Loss (blue=policy, orange=value)",
+            font,
+        )
+        py += chart_h + 12
+
+        # Win rate chart
+        bar_h = 160
+        _draw_stacked_bar(screen, px, py, PANEL_W, bar_h, win_history, font)
+        py += bar_h + 12
+
+        # Latest stats
+        if policy_losses:
+            pl = font.render(f"Policy loss: {policy_losses[-1]:.4f}", True, POLICY_LINE)
+            screen.blit(pl, (px, py))
+            py += 18
+        if value_losses:
+            vl = font.render(f"Value loss:  {value_losses[-1]:.4f}", True, VALUE_LINE)
+            screen.blit(vl, (px, py))
+            py += 18
+        if win_history:
+            p1, p2, dr = win_history[-1]
+            ws = font.render(f"Last iter: P1={p1}  P2={p2}  Draw={dr}", True, TEXT_COLOR)
+            screen.blit(ws, (px, py))
+
+        pygame.display.flip()
+        clock.tick(FPS)
+
+    pygame.quit()
+    _state["running"] = False
+    train_thread.join(timeout=5)
diff --git a/src/main.cpp b/src/main.cpp
index 54c7d7f..53befac 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -121,6 +121,7 @@ bool checkGameEnd();
 void updateThinkingVisuals(int8_t pColor, int8_t column);
 void animateDrop(int col, int player);
 void moveDiscToCol(int startCol, int targetCol, int player, int speed);
+int evaluateBoard(int8_t aiP, int8_t huP);
 int minimax(int depth, int alpha, int beta, bool isMax, int8_t aiP, int8_t huP, int8_t rootCol);
 void performAiMove(int8_t aiP);
 void randomizeDemoPlies();
@@ -265,6 +266,39 @@ int8_t scanBoard() {
     return 0;
 }
 
+int evaluateBoard(int8_t aiP, int8_t huP) {
+    int score = 0;
+
+    // Center column bonus
+    for (int r = 0; r < ROWS; r++) {
+        if (board[3][r] == aiP) score += 3;
+        else if (board[3][r] == huP) score -= 3;
+    }
+
+    // Score a window of 4 cells by piece counts
+    auto scoreWindow = [&](int c, int r, int dc, int dr) -> int {
+        int ai = 0, hu = 0;
+        for (int i = 0; i < 4; i++) {
+            int8_t v = board[c + i * dc][r + i * dr];
+            if (v == aiP) ai++;
+            else if (v == huP) hu++;
+        }
+        if (ai > 0 && hu > 0) return 0;
+        if (ai == 3) return 50;
+        if (ai == 2) return 5;
+        if (hu == 3) return -50;
+        if (hu == 2) return -5;
+        return 0;
+    };
+
+    for (int r = 0; r < 6; r++) for (int c = 0; c < 4; c++) score += scoreWindow(c, r, 1, 0);
+    for (int r = 0; r < 3; r++) for (int c = 0; c < 7; c++) score += scoreWindow(c, r, 0, 1);
+    for (int r = 0; r < 3; r++) for (int c = 0; c < 4; c++) score += scoreWindow(c, r, 1, 1);
+    for (int r = 3; r < 6; r++) for (int c = 0; c < 4; c++) score += scoreWindow(c, r, 1, -1);
+
+    return score;
+}
+
 bool checkGameEnd() {
     winnerPlayer = scanBoard();
     bool won = winnerPlayer != 0;
@@ -331,7 +365,7 @@ int minimax(int depth, int alpha, int beta, bool isMax, int8_t aiP, int8_t huP,
     int8_t win = scanBoard();
     if (win == aiP) return 1000 + depth;
     if (win == huP) return -1000 - depth;
-    if (depth == 0 || isBoardFull()) return 0;
+    if (depth == 0 || isBoardFull()) return evaluateBoard(aiP, huP);
 
     int best = isMax ? -10000 : 10000;
     for (int c : colOrder) {
@@ -356,13 +390,22 @@ void performAiMove(int8_t aiP) {
     int originalPly = currentLookAhead;
     if (gameState == DEMO) currentLookAhead = demoPly[aiP - 1];
 
-    // Phase 1: always take an instant win or block an opponent's win
+    // Phase 1a: check ALL columns for instant AI win
     bool found = false;
     for (int c = 0; c < COLS && !found; c++) {
         int r = getFirstEmptyRow(c);
         if (r != -1) {
-            board[c][r] = aiP; if (scanBoard() == aiP) { board[c][r]=0; bestCol=c; found=true; break; }
-            board[c][r] = huP; if (scanBoard() == huP) { board[c][r]=0; bestCol=c; found=true; break; }
+            board[c][r] = aiP;
+            if (scanBoard() == aiP) { board[c][r] = 0; bestCol = c; found = true; break; }
+            board[c][r] = 0;
+        }
+    }
+    // Phase 1b: check ALL columns for opponent block
+    for (int c = 0; c < COLS && !found; c++) {
+        int r = getFirstEmptyRow(c);
+        if (r != -1) {
+            board[c][r] = huP;
+            if (scanBoard() == huP) { board[c][r] = 0; bestCol = c; found = true; break; }
             board[c][r] = 0;
         }
     }