diff --git a/mixer.py b/mixer.py
new file mode 100644
index 0000000..78cfc8f
--- /dev/null
+++ b/mixer.py
@@ -0,0 +1,236 @@
+# mixer.py
+# ---------------------------------------------------------------------------
+# Mixer module for Amplify
+#
+# Responsibilities:
+#   - Load decoded audio arrays for each timeline item
+#   - Apply any queued ops (scale, loop) to each track
+#   - Stack (sum) all tracks into a single stereo output array
+#   - Optionally normalize the final mix to prevent clipping
+#
+# This module is intentionally kept separate from cli.py and config.py so
+# that the render logic can be tested and extended independently.
+# ---------------------------------------------------------------------------
+
+import numpy as np
+from pathlib import Path
+
+
+# ---------------------------------------------------------------------------
+# Internal helpers
+# ---------------------------------------------------------------------------
+
+def _load_audio(path: str, sample_rate: int, channels: int) -> np.ndarray:
+    """
+    Load an audio file from disk and return it as a float32 numpy array
+    of shape (num_samples, channels), resampled to `sample_rate` if needed.
+
+    Currently a STUB — returns a short block of silence so the rest of the
+    pipeline can be developed and tested before the audio-decode layer is
+    wired up.  Replace the body of this function (and add pydub/soundfile/
+    librosa as a dependency) when the Audio File Input System is ready.
+
+    Args:
+        path:        Absolute or relative path to the audio file.
+        sample_rate: Target sample rate in Hz (e.g. 44100).
+        channels:    Number of output channels (1 = mono, 2 = stereo).
+
+    Returns:
+        np.ndarray of shape (num_samples, channels), dtype float32,
+        values in [-1.0, 1.0].
+    """
+    # TODO: replace with real decode logic once FD-1 library is chosen
+    #       e.g. soundfile.read(path, dtype="float32", always_2d=True)
+    stub_duration_sec = 2
+    num_samples = sample_rate * stub_duration_sec
+    return np.zeros((num_samples, channels), dtype=np.float32)
+
+
+def _apply_scale(audio: np.ndarray, factor: float, preserve_pitch: bool,
+                 sample_rate: int) -> np.ndarray:
+    """
+    Time-scale `audio` by `factor` (0.5–2.0).
+
+    Currently uses simple linear resampling (no pitch preservation).
+    When preserve_pitch=True, a phase-vocoder pass should be applied
+    before the resample — marked as TODO below.
+
+    Args:
+        audio:          Input array (num_samples, channels), float32.
+        factor:         Speed multiplier.  >1.0 speeds up, <1.0 slows down.
+        preserve_pitch: If True, attempt to keep pitch constant (phase vocoder).
+        sample_rate:    Sample rate, needed by pitch-preservation code later.
+
+    Returns:
+        Resampled array with approximately len(audio) / factor samples.
+    """
+    if factor == 1.0:
+        return audio  # nothing to do
+
+    original_len = audio.shape[0]
+    new_len = int(round(original_len / factor))
+
+    # TODO: if preserve_pitch, apply phase-vocoder here before resampling
+    #       (Feature 2.1.10 in the proposal)
+
+    # Simple linear interpolation resample — meets the "minor artifacts
+    # acceptable for MVP" requirement from section 2.1.9
+    old_indices = np.linspace(0, original_len - 1, new_len)
+    left = np.floor(old_indices).astype(int)
+    right = np.clip(left + 1, 0, original_len - 1)
+    frac = (old_indices - left)[:, np.newaxis]   # broadcast over channels
+
+    resampled = audio[left] * (1 - frac) + audio[right] * frac
+    return resampled.astype(np.float32)
+
+
+def _apply_loop(audio: np.ndarray, count: int | None, bpm: float | None,
+                bars: int | None, sample_rate: int) -> np.ndarray:
+    """
+    Repeat `audio` according to either a raw count or a musical meter.
+
+    If `count` is given, the audio is simply tiled that many times.
+    If `bpm` and `bars` are given, the desired loop length in samples is
+    computed from the tempo and time signature, and the audio is tiled then
+    trimmed/padded to that exact length (within 1% as required by 2.1.11).
+
+    Args:
+        audio:       Input array (num_samples, channels), float32.
+        count:       Number of times to repeat, or None.
+        bpm:         Beats per minute, or None.
+        bars:        Number of bars to fill, or None.
+        sample_rate: Sample rate used for BPM math.
+
+    Returns:
+        Looped array, float32.
+    """
+    if count is not None:
+        # Simple tile — stacks the array `count` times along axis 0
+        return np.tile(audio, (count, 1))
+
+    if bpm is not None and bars is not None:
+        # beats_per_bar is hard-coded to 4 for now (4/4 time signature).
+        # TODO: read time_signature from the project config and parse the
+        #       numerator so non-4/4 meters are supported.
+        beats_per_bar = 4
+        seconds_per_beat = 60.0 / bpm
+        target_samples = int(round(bars * beats_per_bar * seconds_per_beat * sample_rate))
+
+        # Tile enough copies to exceed the target, then trim to exact length
+        needed_copies = int(np.ceil(target_samples / max(audio.shape[0], 1)))
+        tiled = np.tile(audio, (max(needed_copies, 1), 1))
+        return tiled[:target_samples]
+
+    # No loop instruction — return audio unchanged
+    return audio
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+def render(config: dict) -> np.ndarray:
+    """
+    Render a full composition defined by `config` into a single float32 array.
+
+    Processing order for each timeline item:
+        1. Load raw audio from disk (or stub)
+        2. Apply ops in order: currently 'scale' and 'loop' are supported
+        3. Offset the track in time if a non-zero 'start' is specified
+        4. Accumulate into the master mix buffer
+
+    After all tracks are summed:
+        5. Normalize the mix if config["mix"]["normalize"] is True
+
+    Args:
+        config: A dict matching the schema produced by config.py (loaded from
+                the project YAML).  Expected top-level keys:
+                  "project"  → sample_rate, channels
+                  "assets"   → list of {id, path}
+                  "timeline" → list of {id, asset, start, gain_db, ops:[]}
+                  "mix"      → {normalize: bool}
+
+    Returns:
+        np.ndarray of shape (total_samples, channels), dtype float32.
+        Returns an empty (0, channels) array if the timeline is empty.
+    """
+    project   = config.get("project", {})
+    sample_rate = int(project.get("sample_rate", 44100))
+    channels    = int(project.get("channels", 2))
+
+    # Build a quick lookup from asset id → file path
+    asset_map = {a["id"]: a["path"] for a in config.get("assets", [])}
+
+    # We'll collect rendered tracks here before summing
+    rendered_tracks: list[np.ndarray] = []
+
+    for item in config.get("timeline", []):
+        asset_id = item.get("asset")
+        if asset_id not in asset_map:
+            # Asset referenced in timeline but not declared in assets block —
+            # skip with a warning rather than crashing (matches 2.1.7 error
+            # handling requirement)
+            print(f"[mixer] Warning: asset '{asset_id}' not found, skipping.")
+            continue
+
+        # --- 1. Load audio ---
+        audio = _load_audio(asset_map[asset_id], sample_rate, channels)
+
+        # --- 2. Apply ops in the order they were queued ---
+        for op in item.get("ops", []):
+            op_type = op.get("type")
+
+            if op_type == "scale":
+                audio = _apply_scale(
+                    audio,
+                    factor=float(op.get("factor", 1.0)),
+                    preserve_pitch=bool(op.get("preserve_pitch", False)),
+                    sample_rate=sample_rate,
+                )
+
+            elif op_type == "loop":
+                audio = _apply_loop(
+                    audio,
+                    count=op.get("count"),
+                    bpm=op.get("bpm"),
+                    bars=op.get("bars"),
+                    sample_rate=sample_rate,
+                )
+
+            # TODO: add more op types here as they are implemented
+            #       e.g. "fade_in", "fade_out", "eq", "reverb"
+
+        # --- 3. Apply gain (convert dB to linear amplitude) ---
+        gain_db = float(item.get("gain_db", 0.0))
+        if gain_db != 0.0:
+            audio = audio * (10.0 ** (gain_db / 20.0))
+
+        # --- 4. Apply start offset (prepend silence) ---
+        start_sec = float(item.get("start", 0.0))
+        if start_sec > 0.0:
+            offset_samples = int(round(start_sec * sample_rate))
+            silence = np.zeros((offset_samples, channels), dtype=np.float32)
+            audio = np.vstack([silence, audio])
+
+        rendered_tracks.append(audio)
+
+    if not rendered_tracks:
+        return np.zeros((0, channels), dtype=np.float32)
+
+    # --- Sum all tracks into a single mix buffer ---
+    # Pad shorter tracks with trailing silence so they all match the longest
+    max_len = max(t.shape[0] for t in rendered_tracks)
+    mix = np.zeros((max_len, channels), dtype=np.float32)
+    for track in rendered_tracks:
+        pad_len = max_len - track.shape[0]
+        if pad_len > 0:
+            track = np.vstack([track, np.zeros((pad_len, channels), dtype=np.float32)])
+        mix += track
+
+    # --- 5. Normalize to prevent clipping ---
+    if config.get("mix", {}).get("normalize", True):
+        peak = np.max(np.abs(mix))
+        if peak > 1.0:
+            mix = mix / peak   # scale down to [-1.0, 1.0]
+
+    return mix
diff --git a/pyproject.toml b/pyproject.toml
index 5b62503..024070c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -10,6 +10,7 @@ requires-python = ">=3.10"
 dependencies = [
     "typer>=0.12",
     "pyyaml>=6.0"
+    "numpy>=1.26"
 ]
 
 [project.scripts]