diff --git a/.opencode/plans/websocket-live-preview.md b/.opencode/plans/websocket-live-preview.md
new file mode 100644
index 0000000..8c96b72
--- /dev/null
+++ b/.opencode/plans/websocket-live-preview.md
@@ -0,0 +1,266 @@
+# Plan: 15+ FPS Live Preview via WebSocket
+
+## Goal
+Achieve **15+ FPS at 640p** live preview during video processing using **WebSocket push** instead of HTTP polling.
+
+---
+
+## Current Architecture Analysis
+
+| Component | Current | Bottleneck |
+|-----------|---------|------------|
+| Frame capture | Every 10th frame (2.5 Hz at 25 FPS) | Too slow for 15 FPS target |
+| Encoding | `cv2.imencode('.jpg', viz)` synchronous, full-res | Blocks pipeline |
+| Transport | HTTP GET `/api/jobs/<id>/frame` every 2s | 2s latency, wasted polls |
+| Frontend | `<img src>` replacement | Flicker, no frame timing control |
+
+**Key finding**: JPEG encoding at 640p = **285-371 FPS** (CPU). Not the bottleneck. The bottleneck is **synchronous encoding + polling architecture**.
+
+---
+
+## Recommended Architecture: WebSocket with Threaded Encoder
+
+```
+┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
+│  Pipeline       │     │  Encoder Thread │     │  Flask-SocketIO │
+│  (detection)    │────▶│  (non-blocking) │────▶│  WebSocket      │
+│  - YOLO detect  │     │  - Downscale    │     │  Server         │
+│  - Track        │     │  - JPEG encode  │     │  - Broadcast    │
+│  - Count        │     │  - Queue push   │     │    to clients   │
+└─────────────────┘     └─────────────────┘     └─────────────────┘
+         │                        │                        │
+         ▼                        ▼                        ▼
+   Full speed            15+ FPS @ 640p            Real-time push
+   processing            to queue                   to browsers
+```
+
+### Why This Design?
+- **Pipeline stays at full speed** — detection/tracking never blocked by encoding
+- **Encoder thread** — produces 15+ FPS at 640p, drops frames if queue full
+- **Flask-SocketIO** — handles WebSocket connections, broadcasting, reconnection
+- **Eventlet/gevent** — async I/O for many concurrent viewers
+
+---
+
+## Implementation Plan
+
+### Phase 1: Dependencies & Server Setup
+
+**Add to `pyproject.toml`:**
+```toml
+[project.optional-dependencies]
+web = [
+    "flask-socketio>=5.3.0",
+    "eventlet>=0.33.0",  # or gevent
+]
+```
+
+**Run:** `pip install -e ".[web]"`
+
+### Phase 2: Pipeline — Non-blocking Frame Producer
+
+**Modify `src/pipeline.py`:**
+- Add `preview_queue: Queue[bytes]` parameter to `run_pipeline()`
+- In main loop: every N frames (configurable), downscale → encode → `queue.put_nowait(jpeg_bytes)`
+- If queue full: drop frame (don't block)
+- Remove `frame_callback` — replaced by queue
+
+```python
+def run_pipeline(
+    ...,
+    preview_queue: "Queue[bytes] | None" = None,
+    preview_every_n: int = 2,      # 25 FPS / 2 = 12.5 FPS → use 1 for 25 FPS
+    preview_max_dim: int = 640,
+    preview_jpeg_quality: int = 70,
+) -> PipelineResult:
+    if preview_queue is not None:
+        preview_interval = max(1, preview_every_n)
+    ...
+    if preview_queue and frame_idx % preview_interval == 0:
+        # Downscale
+        h, w = viz.shape[:2]
+        if max(h, w) > preview_max_dim:
+            scale = preview_max_dim / max(h, w)
+            preview_viz = cv2.resize(viz, (int(w*scale), int(h*scale)))
+        else:
+            preview_viz = viz
+        ok, jpeg = cv2.imencode('.jpg', preview_viz, [cv2.IMWRITE_JPEG_QUALITY, preview_jpeg_quality])
+        if ok:
+            try:
+                preview_queue.put_nowait(jpeg.tobytes())
+            except queue.Full:
+                pass  # Drop frame, don't block
+```
+
+### Phase 3: Job Queue — WebSocket Broadcaster
+
+**Modify `src/job.py`:**
+- Add `preview_queue` per job
+- Start **encoder thread** that reads queue → broadcasts via SocketIO
+- Thread lifecycle tied to job
+
+```python
+import queue
+import threading
+
+class Job:
+    ...
+    preview_queue: "queue.Queue[bytes]" = field(default_factory=lambda: queue.Queue(maxsize=30))
+    _encoder_thread: threading.Thread | None = None
+    _encoder_stop: threading.Event = field(default_factory=threading.Event)
+
+def _start_preview_broadcaster(self, job_id: str, socketio):
+    """Thread: read preview queue, emit via SocketIO."""
+    while not self._encoder_stop.is_set():
+        try:
+            jpeg_bytes = self.preview_queue.get(timeout=0.1)
+            # Emit to job's room
+            socketio.emit('preview_frame', {'frame': base64.b64encode(jpeg_bytes).decode()}, room=job_id)
+        except queue.Empty:
+            continue
+```
+
+### Phase 4: Flask-SocketIO Server
+
+**Modify `app.py`:**
+```python
+from flask_socketio import SocketIO, emit, join_room, leave_room
+
+socketio = SocketIO(app, cors_allowed_origins="*", async_mode='eventlet')
+
+@socketio.on('connect')
+def on_connect():
+    pass
+
+@socketio.on('join_job')
+def on_join_job(data):
+    job_id = data.get('job_id')
+    if job_id:
+        join_room(job_id)
+
+@socketio.on('leave_job')
+def on_leave_job(data):
+    job_id = data.get('job_id')
+    if job_id:
+        leave_room(job_id)
+
+# In _run_job():
+# socketio.start_background_task(_start_preview_broadcaster, job_id, socketio)
+```
+
+**Run with eventlet:**
+```python
+# main.py or app.py
+if __name__ == '__main__':
+    socketio.run(app, host='0.0.0.0', port=9000, debug=False)
+```
+
+### Phase 5: Frontend — WebSocket Consumer
+
+**Replace polling in `static/app.js`:**
+```javascript
+// In initStatusPage():
+var socket = io();
+socket.on('connect', function() {
+    socket.emit('join_job', {job_id: jobId});
+});
+
+socket.on('preview_frame', function(data) {
+    if (livePreviewImg && data.frame) {
+        livePreviewImg.src = 'data:image/jpeg;base64,' + data.frame;
+        if (previewPlaceholder) previewPlaceholder.style.display = 'none';
+        livePreviewImg.style.display = 'block';
+    }
+});
+
+// Cleanup on disconnect/page unload
+socket.on('disconnect', function() {
+    socket.emit('leave_job', {job_id: jobId});
+});
+window.addEventListener('beforeunload', function() {
+    socket.emit('leave_job', {job_id: jobId});
+});
+```
+
+### Phase 6: Configuration & Tuning
+
+| Parameter | Default | Tuning Notes |
+|-----------|---------|--------------|
+| `preview_every_n` | 1 | 1 = every frame (25 FPS max), 2 = 12.5 FPS |
+| `preview_max_dim` | 640 | 480 for faster, 720 for quality |
+| `preview_jpeg_quality` | 70 | 50-80 tradeoff |
+| `queue.maxsize` | 30 | ~2 seconds buffer at 15 FPS |
+
+---
+
+## Expected Performance
+
+| Metric | Current | Target |
+|--------|---------|--------|
+| Preview FPS | ~0.5 (polling) | **15-25** |
+| Latency | 2+ seconds | **<100ms** |
+| Pipeline impact | Blocks every 10th frame | **Zero** |
+| Resolution | Full video res | **640p max** |
+| Bandwidth/frame | ~200-500 KB | **~30-50 KB** |
+
+---
+
+## Files to Modify
+
+| File | Changes |
+|------|---------|
+| `pyproject.toml` | Add `flask-socketio`, `eventlet` to optional deps |
+| `src/pipeline.py` | Add `preview_queue` param, producer logic |
+| `src/job.py` | Add preview queue, encoder thread, SocketIO broadcaster |
+| `app.py` | Initialize SocketIO, add connect/join/leave handlers, run with `socketio.run()` |
+| `static/app.js` | Replace polling with SocketIO client, handle `preview_frame` event |
+| `templates/base.html` | Add SocketIO client script (`/socket.io/socket.io.js`) |
+
+---
+
+## Testing Strategy
+
+1. **Unit test**: `run_pipeline` with mock queue — verify frames enqueued at correct interval
+2. **Integration**: Start job, connect browser, verify 15+ FPS in devtools Network tab
+3. **Load test**: Multiple browser tabs → verify broadcast works
+4. **Stress test**: Long video (10+ min) → verify no memory leaks, queue stays bounded
+
+---
+
+## Risks & Mitigations
+
+| Risk | Mitigation |
+|------|------------|
+| Eventlet monkey-patches stdlib — may break torch/CUDA | Test early; fallback to `gevent` or native `websockets` + asyncio |
+| Queue memory growth | Bounded queue (`maxsize=30`), drop frames when full |
+| Multiple clients | SocketIO rooms — single broadcast to room |
+| Reconnection | SocketIO handles auto-reconnect; client re-joins room on reconnect |
+
+---
+
+## Dependencies Check
+
+- `flask-socketio>=5.3.0` — compatible with Flask 3.x
+- `eventlet>=0.33.0` — works on ARM64 (Jetson)
+- `python-engineio>=4.7.0` — transitive
+
+---
+
+## Rollout
+
+1. Add deps, install
+2. Implement pipeline + job queue changes (backend only)
+3. Test with `python -c "from app import socketio; print('OK')"`
+4. Implement frontend WebSocket client
+5. Test end-to-end
+6. Deploy
+
+---
+
+## Alternative: H.264 Streaming (Future)
+
+If JPEG-over-WebSocket isn't smooth enough:
+- Use `cv2.VideoWriter` with H.264 in encoder thread
+- Stream via HTTP chunked transfer or WebRTC
+- Browser `<video>` tag with `MediaSource` API
+- More complex but hardware-decoded, smoother at high FPS
\ No newline at end of file
diff --git a/app.py b/app.py
index 5e643d3..2ad00f8 100644
--- a/app.py
+++ b/app.py
@@ -10,6 +10,7 @@ from flask import (
     Flask, render_template, request, redirect,
     url_for, send_file, jsonify,
 )
+from flask_socketio import SocketIO, join_room, leave_room
 from werkzeug.utils import secure_filename
 
 from pathlib import Path
@@ -24,6 +25,8 @@ app = Flask(__name__, template_folder="templates", static_folder="static")
 app.config["SECRET_KEY"] = os.getenv("SECRET_KEY", "change-me")
 app.config["MAX_CONTENT_LENGTH"] = 2 * 1024 * 1024 * 1024  # 2GB
 
+socketio = SocketIO(app, cors_allowed_origins="*", async_mode='threading')
+
 MODELS_DIR = os.getenv("MODELS_DIR", "./models")
 UPLOAD_DIR = os.getenv("UPLOAD_DIR", "./uploads")
 OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./output")
@@ -36,6 +39,30 @@ job_queue = JobQueue(output_dir=OUTPUT_DIR)
 VIDEO_EXTENSIONS = ('.mp4', '.avi', '.mkv', '.mov', '.webm')
 
 
+@socketio.on('connect')
+def on_connect():
+    pass
+
+
+@socketio.on('disconnect')
+def on_disconnect():
+    pass
+
+
+@socketio.on('join_job')
+def on_join_job(data):
+    job_id = data.get('job_id')
+    if job_id:
+        join_room(job_id)
+
+
+@socketio.on('leave_job')
+def on_leave_job(data):
+    job_id = data.get('job_id')
+    if job_id:
+        leave_room(job_id)
+
+
 def _parse_model_configs(form):
     """Parse model selection from request form.
 
@@ -128,6 +155,7 @@ def upload():
         model_configs=model_configs,
         class_filters=class_filters,
     )
+    job.start_preview_broadcaster(job.job_id, socketio)
 
     return redirect(url_for("status", job_id=job.job_id))
 
@@ -148,6 +176,7 @@ def upload_reuse():
         model_configs=model_configs,
         class_filters=class_filters,
     )
+    job.start_preview_broadcaster(job.job_id, socketio)
 
     return redirect(url_for("status", job_id=job.job_id))
 
@@ -337,7 +366,7 @@ def main():
     debug = os.getenv("FLASK_DEBUG", "false").lower() == "true"
 
     print(f"Feedmill Recounter web UI: http://{host}:{port}")
-    app.run(host=host, port=port, debug=debug)
+    socketio.run(app, host=host, port=port, debug=debug, allow_unsafe_werkzeug=True)
 
 
 if __name__ == "__main__":
diff --git a/pyproject.toml b/pyproject.toml
index 2a667d7..3576fe4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -18,6 +18,7 @@ dependencies = [
 
 [project.optional-dependencies]
 dev = ["pytest"]
+web = ["flask-socketio>=5.3.0", "eventlet>=0.33.0"]
 
 [project.scripts]
 recounter = "cli:main"
diff --git a/src/job.py b/src/job.py
index ab8b94b..5222beb 100644
--- a/src/job.py
+++ b/src/job.py
@@ -3,7 +3,9 @@
 
 from __future__ import annotations
 
+import base64
 import os
+import queue
 import threading
 import time
 import uuid
@@ -55,6 +57,31 @@ class Job:
     created_at: float = field(default_factory=time.time)
     completed_at: float | None = None
     latest_frame: bytes | None = None
+    preview_queue: queue.Queue = field(default_factory=lambda: queue.Queue(maxsize=30))
+    _preview_thread: threading.Thread | None = None
+    _preview_stop: threading.Event = field(default_factory=threading.Event)
+    _job_lock: threading.Lock = field(default_factory=threading.Lock)
+
+    def start_preview_broadcaster(self, job_id: str, socketio) -> None:
+        """Start background thread that reads preview_queue and emits via SocketIO."""
+        def _broadcaster():
+            while not self._preview_stop.is_set():
+                try:
+                    jpeg_bytes = self.preview_queue.get(timeout=0.2)
+                    with self._job_lock:
+                        self.latest_frame = jpeg_bytes
+                    encoded = base64.b64encode(jpeg_bytes).decode('ascii')
+                    socketio.emit('preview_frame', {'frame': encoded}, room=job_id)
+                except queue.Empty:
+                    continue
+                except Exception:
+                    continue
+
+        self._preview_thread = threading.Thread(target=_broadcaster, daemon=True)
+        self._preview_thread.start()
+
+    def stop_preview_broadcaster(self) -> None:
+        self._preview_stop.set()
 
 
 class JobQueue:
@@ -154,16 +181,13 @@ class JobQueue:
                     f"{model_cfg.stem}_annotated.mp4",
                 )
 
-                def _on_frame(jpeg_bytes: bytes) -> None:
-                    with self._lock:
-                        job.latest_frame = jpeg_bytes
-
                 result: PipelineResult = run_pipeline(
                     video_path=job.video_path,
                     model_config=model_cfg,
                     output_path=output_path,
                     class_filter=class_filter,
-                    frame_callback=_on_frame,
+                    preview_queue=job.preview_queue,
+                    preview_every_n=2,
                     cancel_check=lambda: job.status == JobStatus.CANCELLED,
                 )
 
diff --git a/src/pipeline.py b/src/pipeline.py
index 95b1983..0db4913 100644
--- a/src/pipeline.py
+++ b/src/pipeline.py
@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import queue
 import time
 from collections.abc import Callable
 from dataclasses import dataclass
@@ -63,6 +64,10 @@ def run_pipeline(
     progress_callback: Callable[[int, int], None] | None = None,
     frame_callback: Callable[[bytes], None] | None = None,
     cancel_check: Callable[[], bool] | None = None,
+    preview_queue: queue.Queue[bytes] | None = None,
+    preview_every_n: int = 2,
+    preview_max_dim: int = 640,
+    preview_jpeg_quality: int = 70,
 ) -> PipelineResult:
     """Process a video file through the counting pipeline.
 
@@ -75,8 +80,12 @@ def run_pipeline(
         truck_conf: Truck detection confidence threshold (default: 0.5).
         truck_det_interval: Run truck detection every N frames.
         progress_callback: Optional fn(frame_idx, total_frames) called per frame.
-        frame_callback: Optional fn(jpeg_bytes) called every 10th frame.
+        frame_callback: Optional fn(jpeg_bytes) called every 10th frame (fallback).
         cancel_check: Optional fn() returning True to abort processing.
+        preview_queue: Optional queue for live preview frames (JPEG bytes).
+        preview_every_n: Push a preview frame every N frames (default 2).
+        preview_max_dim: Max dimension (w or h) for preview frames.
+        preview_jpeg_quality: JPEG quality for preview frames (1-100).
 
     Returns:
         PipelineResult with counting summary.
@@ -223,7 +232,20 @@ def run_pipeline(
 
             writer.write_frame(viz)
 
-            if frame_callback and frame_idx % 10 == 0:
+            if preview_queue is not None and frame_idx % max(1, preview_every_n) == 0:
+                h, w = viz.shape[:2]
+                if max(h, w) > preview_max_dim:
+                    scale = preview_max_dim / max(h, w)
+                    preview_viz = cv2.resize(viz, (int(w * scale), int(h * scale)))
+                else:
+                    preview_viz = viz
+                ok, jpeg = cv2.imencode('.jpg', preview_viz, [cv2.IMWRITE_JPEG_QUALITY, preview_jpeg_quality])
+                if ok:
+                    try:
+                        preview_queue.put_nowait(jpeg.tobytes())
+                    except Exception:
+                        pass
+            elif frame_callback and frame_idx % 10 == 0:
                 ok, jpeg = cv2.imencode('.jpg', viz)
                 if ok:
                     frame_callback(jpeg.tobytes())
diff --git a/static/app.js b/static/app.js
index b7ee3e3..8a99869 100644
--- a/static/app.js
+++ b/static/app.js
@@ -352,13 +352,30 @@ function initStatusPage(jobId, initialStatus) {
   var cancelBtn = document.getElementById('cancel-btn');
   var terminalStates = ['COMPLETED', 'FAILED', 'CANCELLED'];
   var pollInterval = null;
-  var previewInterval = null;
   var consecutiveFailures = 0;
   var connectionLostIndicator = null;
   var lastStatus = initialStatus;
 
   if (!progressFill) return;
 
+  /* --- WebSocket live preview --- */
+  var socket = io();
+  socket.on('connect', function() {
+    socket.emit('join_job', {job_id: jobId});
+  });
+
+  socket.on('preview_frame', function(data) {
+    if (livePreviewImg && data.frame) {
+      livePreviewImg.src = 'data:image/jpeg;base64,' + data.frame;
+      if (previewPlaceholder) previewPlaceholder.style.display = 'none';
+      livePreviewImg.style.display = 'block';
+    }
+  });
+
+  window.addEventListener('beforeunload', function() {
+    socket.emit('leave_job', {job_id: jobId});
+  });
+
   /* --- Cancel button --- */
   if (cancelBtn) {
     cancelBtn.addEventListener('click', function() {
@@ -402,15 +419,8 @@ function initStatusPage(jobId, initialStatus) {
           progressRegion.setAttribute('aria-busy', data.status === 'RUNNING' || data.status === 'PENDING' ? 'true' : 'false');
         }
 
-        if (data.status === 'RUNNING' || data.status === 'PENDING') {
-          updateLivePreview(jobId);
-        }
-
         if (terminalStates.indexOf(data.status) !== -1) {
           stopPolling();
-          if (data.status !== 'COMPLETED') {
-            updateLivePreview(jobId);
-          }
         }
       })
       .catch(function () {
@@ -460,20 +470,6 @@ function initStatusPage(jobId, initialStatus) {
     modelName.textContent = name || '—';
   }
 
-  /* --- Fetch and display latest annotated frame --- */
-  function updateLivePreview(jid) {
-    if (!livePreviewImg) return;
-    livePreviewImg.src = '/api/jobs/' + jid + '/frame?t=' + Date.now();
-    livePreviewImg.onload = function () {
-      if (previewPlaceholder) previewPlaceholder.style.display = 'none';
-      livePreviewImg.style.display = 'block';
-    };
-    livePreviewImg.onerror = function () {
-      livePreviewImg.style.display = 'none';
-      if (previewPlaceholder) previewPlaceholder.style.display = '';
-    };
-  }
-
   /* --- Render result cards --- */
   function updateResults(results) {
     if (!resultsGrid) return;
@@ -533,14 +529,10 @@ function initStatusPage(jobId, initialStatus) {
   /* --- Start / stop polling --- */
   function startPolling() {
     pollInterval = setInterval(pollJob, 2000);
-    previewInterval = setInterval(function () {
-      updateLivePreview(jobId);
-    }, 2000);
   }
 
   function stopPolling() {
     if (pollInterval) { clearInterval(pollInterval); pollInterval = null; }
-    if (previewInterval) { clearInterval(previewInterval); previewInterval = null; }
   }
 
   /* --- Boot --- */
@@ -558,8 +550,5 @@ function initStatusPage(jobId, initialStatus) {
   }
   if (terminalStates.indexOf(initialStatus) === -1) {
     startPolling();
-    updateLivePreview(jobId);
-  } else {
-    updateLivePreview(jobId);
   }
 }
diff --git a/templates/base.html b/templates/base.html
index 7601732..f557181 100644
--- a/templates/base.html
+++ b/templates/base.html
@@ -35,5 +35,6 @@
         </div>
     </footer>
     {% block scripts %}{% endblock %}
+    <script src="https://cdn.socket.io/4.7.5/socket.io.min.js"></script>
 </body>
 </html>