diff --git a/opengait/demo/__init__.py b/opengait/demo/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/opengait/demo/__main__.py b/opengait/demo/__main__.py
new file mode 100644
index 0000000..2590b2a
--- /dev/null
+++ b/opengait/demo/__main__.py
@@ -0,0 +1,7 @@
+from __future__ import annotations
+
+from .pipeline import main
+
+
+if __name__ == "__main__":
+    main()
diff --git a/opengait/demo/input.py b/opengait/demo/input.py
new file mode 100644
index 0000000..a96f540
--- /dev/null
+++ b/opengait/demo/input.py
@@ -0,0 +1,203 @@
+"""
+Input adapters for OpenGait demo.
+
+Provides generator-based interfaces for video sources:
+- OpenCV (video files, cameras)
+- cv-mmap (shared memory streams)
+"""
+
+from collections.abc import AsyncIterator, Generator, Iterable
+from typing import TYPE_CHECKING, Protocol, cast
+import logging
+
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+# Type alias for frame stream: (frame_array, metadata_dict)
+FrameStream = Iterable[tuple[np.ndarray, dict[str, object]]]
+
+if TYPE_CHECKING:
+    # Protocol for cv-mmap metadata to avoid direct import
+    class _FrameMetadata(Protocol):
+        frame_count: int
+        timestamp_ns: int
+
+    # Protocol for cv-mmap client
+    class _CvMmapClient(Protocol):
+        def __aiter__(self) -> AsyncIterator[tuple[np.ndarray, _FrameMetadata]]: ...
+
+
+def opencv_source(
+    path: str | int, max_frames: int | None = None
+) -> Generator[tuple[np.ndarray, dict[str, object]], None, None]:
+    """
+    Generator that yields frames from an OpenCV video source.
+
+    Parameters
+    ----------
+    path : str | int
+        Video file path or camera index (e.g., 0 for default camera)
+    max_frames : int | None, optional
+        Maximum number of frames to yield. None means unlimited.
+
+    Yields
+    ------
+    tuple[np.ndarray, dict[str, object]]
+        (frame_array, metadata_dict) where metadata includes:
+        - frame_count: frame index (0-based)
+        - timestamp_ns: monotonic timestamp in nanoseconds (if available)
+        - source: the path/int provided
+    """
+    import time
+
+    import cv2
+
+    cap = cv2.VideoCapture(path)
+    if not cap.isOpened():
+        raise RuntimeError(f"Failed to open video source: {path}")
+
+    frame_idx = 0
+    try:
+        while max_frames is None or frame_idx < max_frames:
+            ret, frame = cap.read()
+            if not ret:
+                # End of stream
+                break
+
+            # Get timestamp if available (some backends support this)
+            timestamp_ns = time.monotonic_ns()
+
+            metadata: dict[str, object] = {
+                "frame_count": frame_idx,
+                "timestamp_ns": timestamp_ns,
+                "source": path,
+            }
+
+            yield frame, metadata
+            frame_idx += 1
+
+    finally:
+        cap.release()
+        logger.debug(f"OpenCV source closed: {path}")
+
+
+def cvmmap_source(
+    name: str, max_frames: int | None = None
+) -> Generator[tuple[np.ndarray, dict[str, object]], None, None]:
+    """
+    Generator that yields frames from a cv-mmap shared memory stream.
+
+    Bridges async cv-mmap client to synchronous generator using asyncio.run().
+
+    Parameters
+    ----------
+    name : str
+        Base name of the cv-mmap source (e.g., "default")
+    max_frames : int | None, optional
+        Maximum number of frames to yield. None means unlimited.
+
+    Yields
+    ------
+    tuple[np.ndarray, dict[str, object]]
+        (frame_array, metadata_dict) where metadata includes:
+        - frame_count: frame index from cv-mmap
+        - timestamp_ns: timestamp in nanoseconds from cv-mmap
+        - source: the cv-mmap name
+
+    Raises
+    ------
+    ImportError
+        If cvmmap package is not available
+    RuntimeError
+        If cv-mmap stream disconnects or errors
+    """
+    import asyncio
+
+    # Import cvmmap only when function is called
+    # Use try/except for runtime import check
+    try:
+
+        from cvmmap import CvMmapClient as _CvMmapClientReal  # pyright: ignore[reportMissingTypeStubs]
+    except ImportError as e:
+        raise ImportError(
+            "cvmmap package is required for cv-mmap sources. "
+            + "Install from: https://github.com/crosstyan/cv-mmap"
+        ) from e
+
+    # Cast to protocol type for type checking
+    client: _CvMmapClient = cast("_CvMmapClient", _CvMmapClientReal(name))
+    frame_count = 0
+
+    async def _async_generator() -> AsyncIterator[tuple[np.ndarray, _FrameMetadata]]:
+        """Async generator wrapper."""
+        async for frame, meta in client:
+            yield frame, meta
+
+    # Bridge async to sync using asyncio.run()
+    # We process frames one at a time to keep it simple and robust
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
+
+    try:
+        agen = _async_generator().__aiter__()
+
+        while max_frames is None or frame_count < max_frames:
+            try:
+                frame, meta = loop.run_until_complete(agen.__anext__())
+            except StopAsyncIteration:
+                break
+
+            metadata: dict[str, object] = {
+                "frame_count": meta.frame_count,
+                "timestamp_ns": meta.timestamp_ns,
+                "source": f"cvmmap://{name}",
+            }
+
+            yield frame, metadata
+            frame_count += 1
+
+    finally:
+        loop.close()
+        logger.debug(f"cv-mmap source closed: {name}")
+
+
+def create_source(source: str, max_frames: int | None = None) -> FrameStream:
+    """
+    Factory function to create a frame source from a string specification.
+
+    Parameters
+    ----------
+    source : str
+        Source specification:
+        - '0', '1', etc. -> Camera index (OpenCV)
+        - 'cvmmap://name' -> cv-mmap shared memory stream
+        - Any other string -> Video file path (OpenCV)
+    max_frames : int | None, optional
+        Maximum number of frames to yield. None means unlimited.
+
+    Returns
+    -------
+    FrameStream
+        Generator yielding (frame, metadata) tuples
+
+    Examples
+    --------
+    >>> for frame, meta in create_source('0'):  # Camera 0
+    ...     process(frame)
+    >>> for frame, meta in create_source('cvmmap://default'):  # cv-mmap
+    ...     process(frame)
+    >>> for frame, meta in create_source('/path/to/video.mp4'):
+    ...     process(frame)
+    """
+    # Check for cv-mmap protocol
+    if source.startswith("cvmmap://"):
+        name = source[len("cvmmap://") :]
+        return cvmmap_source(name, max_frames)
+
+    # Check for camera index (single digit string)
+    if source.isdigit():
+        return opencv_source(int(source), max_frames)
+
+    # Otherwise treat as file path
+    return opencv_source(source, max_frames)
diff --git a/opengait/demo/output.py b/opengait/demo/output.py
new file mode 100644
index 0000000..5c9180e
--- /dev/null
+++ b/opengait/demo/output.py
@@ -0,0 +1,368 @@
+"""
+Output publishers for OpenGait demo results.
+
+Provides pluggable result publishing:
+- ConsolePublisher: JSONL to stdout
+- NatsPublisher: NATS message broker integration
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import sys
+import threading
+import time
+from typing import TYPE_CHECKING, Protocol, TextIO, cast, runtime_checkable
+
+if TYPE_CHECKING:
+    from types import TracebackType
+
+logger = logging.getLogger(__name__)
+
+
+@runtime_checkable
+class ResultPublisher(Protocol):
+    """Protocol for result publishers."""
+
+    def publish(self, result: dict[str, object]) -> None:
+        """
+        Publish a result dictionary.
+
+        Parameters
+        ----------
+        result : dict[str, object]
+            Result data with keys: frame, track_id, label, confidence, window, timestamp_ns
+        """
+        ...
+
+
+class ConsolePublisher:
+    """Publisher that outputs JSON Lines to stdout."""
+
+    _output: TextIO
+
+    def __init__(self, output: TextIO = sys.stdout) -> None:
+        """
+        Initialize console publisher.
+
+        Parameters
+        ----------
+        output : TextIO
+            File-like object to write to (default: sys.stdout)
+        """
+        self._output = output
+
+    def publish(self, result: dict[str, object]) -> None:
+        """
+        Publish result as JSON line.
+
+        Parameters
+        ----------
+        result : dict[str, object]
+            Result data with keys: frame, track_id, label, confidence, window, timestamp_ns
+        """
+        try:
+            json_line = json.dumps(result, ensure_ascii=False, default=str)
+            _ = self._output.write(json_line + "\n")
+            self._output.flush()
+        except Exception as e:
+            logger.warning(f"Failed to publish to console: {e}")
+
+    def close(self) -> None:
+        """Close the publisher (no-op for console)."""
+        pass
+
+    def __enter__(self) -> ConsolePublisher:
+        """Context manager entry."""
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
+        """Context manager exit."""
+        self.close()
+
+
+class _NatsClient(Protocol):
+    """Protocol for connected NATS client."""
+
+    async def publish(self, subject: str, payload: bytes) -> object: ...
+
+    async def close(self) -> object: ...
+
+    async def flush(self) -> object: ...
+
+
+class NatsPublisher:
+    """
+    Publisher that sends results to NATS message broker.
+
+    This is a sync-friendly wrapper around the async nats-py client.
+    Uses a background thread with dedicated event loop to bridge sync
+    publish calls to async NATS operations, making it safe to use in
+    both sync and async contexts.
+    """
+
+    _nats_url: str
+    _subject: str
+    _nc: _NatsClient | None
+    _connected: bool
+    _loop: asyncio.AbstractEventLoop | None
+    _thread: threading.Thread | None
+    _lock: threading.Lock
+
+    def __init__(self, nats_url: str, subject: str = "scoliosis.result") -> None:
+        """
+        Initialize NATS publisher.
+
+        Parameters
+        ----------
+        nats_url : str
+            NATS server URL (e.g., "nats://localhost:4222")
+        subject : str
+            NATS subject to publish to (default: "scoliosis.result")
+        """
+        self._nats_url = nats_url
+        self._subject = subject
+        self._nc = None
+        self._connected = False
+        self._loop = None
+        self._thread = None
+        self._lock = threading.Lock()
+
+    def _start_background_loop(self) -> bool:
+        """
+        Start background thread with event loop for async operations.
+
+        Returns
+        -------
+        bool
+            True if loop is running, False otherwise
+        """
+        with self._lock:
+            if self._loop is not None and self._loop.is_running():
+                return True
+
+            try:
+                loop = asyncio.new_event_loop()
+                self._loop = loop
+
+                def run_loop() -> None:
+                    asyncio.set_event_loop(loop)
+                    loop.run_forever()
+
+                self._thread = threading.Thread(target=run_loop, daemon=True)
+                self._thread.start()
+                return True
+            except Exception as e:
+                logger.warning(f"Failed to start background event loop: {e}")
+                return False
+
+    def _stop_background_loop(self) -> None:
+        """Stop the background event loop and thread."""
+        with self._lock:
+            if self._loop is not None and self._loop.is_running():
+                _ = self._loop.call_soon_threadsafe(self._loop.stop)
+            if self._thread is not None and self._thread.is_alive():
+                self._thread.join(timeout=2.0)
+            self._loop = None
+            self._thread = None
+
+    def _ensure_connected(self) -> bool:
+        """
+        Ensure connection to NATS server.
+
+        Returns
+        -------
+        bool
+            True if connected, False otherwise
+        """
+        with self._lock:
+            if self._connected and self._nc is not None:
+                return True
+
+        if not self._start_background_loop():
+            return False
+
+        try:
+            import nats
+
+            async def _connect() -> _NatsClient:
+                nc = await nats.connect(self._nats_url)  # pyright: ignore[reportUnknownMemberType]
+                return cast(_NatsClient, nc)
+
+            # Run connection in background loop
+            future = asyncio.run_coroutine_threadsafe(
+                _connect(),
+                self._loop,  # pyright: ignore[reportArgumentType]
+            )
+            self._nc = future.result(timeout=10.0)
+            self._connected = True
+            logger.info(f"Connected to NATS at {self._nats_url}")
+            return True
+        except ImportError:
+            logger.warning(
+                "nats-py package not installed. Install with: pip install nats-py"
+            )
+            return False
+        except Exception as e:
+            logger.warning(f"Failed to connect to NATS at {self._nats_url}: {e}")
+            return False
+
+    def publish(self, result: dict[str, object]) -> None:
+        """
+        Publish result to NATS subject.
+
+        Parameters
+        ----------
+        result : dict[str, object]
+            Result data with keys: frame, track_id, label, confidence, window, timestamp_ns
+        """
+        if not self._ensure_connected():
+            # Graceful degradation: log warning but don't crash
+            logger.debug(
+                f"NATS unavailable, dropping result: {result.get('track_id', 'unknown')}"
+            )
+            return
+
+        try:
+
+            async def _publish() -> None:
+                if self._nc is not None:
+                    payload = json.dumps(
+                        result, ensure_ascii=False, default=str
+                    ).encode("utf-8")
+                    _ = await self._nc.publish(self._subject, payload)
+                    _ = await self._nc.flush()
+            # Run publish in background loop
+            future = asyncio.run_coroutine_threadsafe(
+                _publish(),
+                self._loop,  # pyright: ignore[reportArgumentType]
+            )
+            future.result(timeout=5.0)  # Wait for publish to complete
+        except Exception as e:
+            logger.warning(f"Failed to publish to NATS: {e}")
+            self._connected = False  # Mark for reconnection on next publish
+
+    def close(self) -> None:
+        """Close NATS connection."""
+        with self._lock:
+            if self._nc is not None and self._connected and self._loop is not None:
+                try:
+
+                    async def _close() -> None:
+                        if self._nc is not None:
+                            _ = await self._nc.close()
+
+                    future = asyncio.run_coroutine_threadsafe(
+                        _close(),
+                        self._loop,
+                    )
+                    future.result(timeout=5.0)
+                except Exception as e:
+                    logger.debug(f"Error closing NATS connection: {e}")
+                finally:
+                    self._nc = None
+                    self._connected = False
+
+        self._stop_background_loop()
+
+    def __enter__(self) -> NatsPublisher:
+        """Context manager entry."""
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_val: BaseException | None,
+        exc_tb: TracebackType | None,
+    ) -> None:
+        """Context manager exit."""
+        self.close()
+
+
+def create_publisher(
+    nats_url: str | None,
+    subject: str = "scoliosis.result",
+) -> ResultPublisher:
+    """
+    Factory function to create appropriate publisher.
+
+    Parameters
+    ----------
+    nats_url : str | None
+        NATS server URL. If None or empty, returns ConsolePublisher.
+    subject : str
+        NATS subject to publish to (default: "scoliosis.result")
+
+    Returns
+    -------
+    ResultPublisher
+        NatsPublisher if nats_url provided, otherwise ConsolePublisher
+
+    Examples
+    --------
+    >>> # Console output (default)
+    >>> pub = create_publisher(None)
+    >>> pub.publish({"frame": 1, "track_id": 42, "label": "normal", "confidence": 0.95, "window": 30, "timestamp_ns": 1234567890})
+    >>>
+    >>> # NATS output
+    >>> pub = create_publisher("nats://localhost:4222")
+    >>> pub.publish({"frame": 1, "track_id": 42, "label": "normal", "confidence": 0.95, "window": 30, "timestamp_ns": 1234567890})
+    >>>
+    >>> # Context manager usage
+    >>> with create_publisher("nats://localhost:4222") as pub:
+    ...     pub.publish(result)
+    """
+    if nats_url:
+        return NatsPublisher(nats_url, subject)
+    return ConsolePublisher()
+
+
+def create_result(
+    frame: int,
+    track_id: int,
+    label: str,
+    confidence: float,
+    window: int | tuple[int, int],
+    timestamp_ns: int | None = None,
+) -> dict[str, object]:
+    """
+    Create a standardized result dictionary.
+
+    Parameters
+    ----------
+    frame : int
+        Frame number
+    track_id : int
+        Track/person identifier
+    label : str
+        Classification label (e.g., "normal", "scoliosis")
+    confidence : float
+        Confidence score (0.0 to 1.0)
+    window : int | tuple[int, int]
+        Frame window as int (end frame) or tuple [start, end] that produced this result
+        Frame window [start, end] that produced this result
+    timestamp_ns : int | None
+        Timestamp in nanoseconds. If None, uses current time.
+
+    Returns
+    -------
+    dict[str, object]
+        Standardized result dictionary
+    """
+    return {
+        "frame": frame,
+        "track_id": track_id,
+        "label": label,
+        "confidence": confidence,
+        "window": window if isinstance(window, int) else window[1],
+        "timestamp_ns": timestamp_ns
+        if timestamp_ns is not None
+        else time.monotonic_ns(),
+    }
diff --git a/opengait/demo/pipeline.py b/opengait/demo/pipeline.py
new file mode 100644
index 0000000..28d5a61
--- /dev/null
+++ b/opengait/demo/pipeline.py
@@ -0,0 +1,325 @@
+from __future__ import annotations
+
+from collections.abc import Callable
+from contextlib import suppress
+import logging
+from pathlib import Path
+import time
+from typing import Protocol, cast
+
+from beartype import beartype
+import click
+import jaxtyping
+from jaxtyping import Float, UInt8
+import numpy as np
+from numpy import ndarray
+from numpy.typing import NDArray
+from ultralytics.models.yolo.model import YOLO
+
+from .input import FrameStream, create_source
+from .output import ResultPublisher, create_publisher, create_result
+from .preprocess import frame_to_person_mask, mask_to_silhouette
+from .sconet_demo import ScoNetDemo
+from .window import SilhouetteWindow, select_person
+
+logger = logging.getLogger(__name__)
+
+JaxtypedDecorator = Callable[[Callable[..., object]], Callable[..., object]]
+JaxtypedFactory = Callable[..., JaxtypedDecorator]
+jaxtyped = cast(JaxtypedFactory, jaxtyping.jaxtyped)
+
+
+class _BoxesLike(Protocol):
+    @property
+    def xyxy(self) -> NDArray[np.float32] | object: ...
+
+    @property
+    def id(self) -> NDArray[np.int64] | object | None: ...
+
+
+class _MasksLike(Protocol):
+    @property
+    def data(self) -> NDArray[np.float32] | object: ...
+
+
+class _DetectionResultsLike(Protocol):
+    @property
+    def boxes(self) -> _BoxesLike: ...
+
+    @property
+    def masks(self) -> _MasksLike: ...
+
+
+class _TrackCallable(Protocol):
+    def __call__(
+        self,
+        source: object,
+        *,
+        persist: bool = True,
+        verbose: bool = False,
+        device: str | None = None,
+        classes: list[int] | None = None,
+    ) -> object: ...
+
+
+class ScoliosisPipeline:
+    _detector: object
+    _source: FrameStream
+    _window: SilhouetteWindow
+    _publisher: ResultPublisher
+    _classifier: ScoNetDemo
+    _device: str
+    _closed: bool
+
+    def __init__(
+        self,
+        *,
+        source: str,
+        checkpoint: str,
+        config: str,
+        device: str,
+        yolo_model: str,
+        window: int,
+        stride: int,
+        nats_url: str | None,
+        nats_subject: str,
+        max_frames: int | None,
+    ) -> None:
+        self._detector = YOLO(yolo_model)
+        self._source = create_source(source, max_frames=max_frames)
+        self._window = SilhouetteWindow(window_size=window, stride=stride)
+        self._publisher = create_publisher(nats_url=nats_url, subject=nats_subject)
+        self._classifier = ScoNetDemo(
+            cfg_path=config,
+            checkpoint_path=checkpoint,
+            device=device,
+        )
+        self._device = device
+        self._closed = False
+
+    @staticmethod
+    def _extract_int(meta: dict[str, object], key: str, fallback: int) -> int:
+        value = meta.get(key)
+        if isinstance(value, int):
+            return value
+        return fallback
+
+    @staticmethod
+    def _extract_timestamp(meta: dict[str, object]) -> int:
+        value = meta.get("timestamp_ns")
+        if isinstance(value, int):
+            return value
+        return time.monotonic_ns()
+
+    @staticmethod
+    def _to_mask_u8(mask: ndarray) -> UInt8[ndarray, "h w"]:
+        binary = np.where(np.asarray(mask) > 0.5, np.uint8(255), np.uint8(0)).astype(
+            np.uint8
+        )
+        return cast(UInt8[ndarray, "h w"], binary)
+
+    def _first_result(self, detections: object) -> _DetectionResultsLike | None:
+        if isinstance(detections, list):
+            return cast(_DetectionResultsLike, detections[0]) if detections else None
+        if isinstance(detections, tuple):
+            return cast(_DetectionResultsLike, detections[0]) if detections else None
+        return cast(_DetectionResultsLike, detections)
+
+    def _select_silhouette(
+        self,
+        result: _DetectionResultsLike,
+    ) -> tuple[Float[ndarray, "64 44"], int] | None:
+        selected = select_person(result)
+        if selected is not None:
+            mask_raw, bbox, track_id = selected
+            silhouette = cast(
+                Float[ndarray, "64 44"] | None,
+                mask_to_silhouette(self._to_mask_u8(mask_raw), bbox),
+            )
+            if silhouette is not None:
+                return silhouette, int(track_id)
+
+        fallback = cast(
+            tuple[UInt8[ndarray, "h w"], tuple[int, int, int, int]] | None,
+            frame_to_person_mask(result),
+        )
+        if fallback is None:
+            return None
+
+        mask_u8, bbox = fallback
+        silhouette = cast(
+            Float[ndarray, "64 44"] | None,
+            mask_to_silhouette(mask_u8, bbox),
+        )
+        if silhouette is None:
+            return None
+        return silhouette, 0
+
+    @jaxtyped(typechecker=beartype)
+    def process_frame(
+        self,
+        frame: UInt8[ndarray, "h w c"],
+        metadata: dict[str, object],
+    ) -> dict[str, object] | None:
+        frame_idx = self._extract_int(metadata, "frame_count", fallback=0)
+        timestamp_ns = self._extract_timestamp(metadata)
+
+        track_fn_obj = getattr(self._detector, "track", None)
+        if not callable(track_fn_obj):
+            raise RuntimeError("YOLO detector does not expose a callable track()")
+
+        track_fn = cast(_TrackCallable, track_fn_obj)
+        detections = track_fn(
+            frame,
+            persist=True,
+            verbose=False,
+            device=self._device,
+            classes=[0],
+        )
+        first = self._first_result(detections)
+        if first is None:
+            return None
+
+        selected = self._select_silhouette(first)
+        if selected is None:
+            return None
+
+        silhouette, track_id = selected
+        self._window.push(silhouette, frame_idx=frame_idx, track_id=track_id)
+
+        if not self._window.should_classify():
+            return None
+
+        window_tensor = self._window.get_tensor(device=self._device)
+        label, confidence = cast(
+            tuple[str, float],
+            self._classifier.predict(window_tensor),
+        )
+        self._window.mark_classified()
+
+        window_start = frame_idx - self._window.window_size + 1
+        result = create_result(
+            frame=frame_idx,
+            track_id=track_id,
+            label=label,
+            confidence=float(confidence),
+            window=(max(0, window_start), frame_idx),
+            timestamp_ns=timestamp_ns,
+        )
+        self._publisher.publish(result)
+        return result
+
+    def run(self) -> int:
+        frame_count = 0
+        start_time = time.perf_counter()
+        try:
+            for item in self._source:
+                frame, metadata = item
+                frame_u8 = np.asarray(frame, dtype=np.uint8)
+                frame_idx = self._extract_int(metadata, "frame_count", fallback=0)
+                frame_count += 1
+                try:
+                    _ = self.process_frame(frame_u8, metadata)
+                except Exception as frame_error:
+                    logger.warning(
+                        "Skipping frame %d due to processing error: %s",
+                        frame_idx,
+                        frame_error,
+                    )
+                if frame_count % 100 == 0:
+                    elapsed = time.perf_counter() - start_time
+                    fps = frame_count / elapsed if elapsed > 0 else 0.0
+                    logger.info("Processed %d frames (%.2f FPS)", frame_count, fps)
+            return 0
+        except KeyboardInterrupt:
+            logger.info("Interrupted by user, shutting down cleanly.")
+            return 130
+        finally:
+            self.close()
+
+    def close(self) -> None:
+        if self._closed:
+            return
+        close_fn = getattr(self._publisher, "close", None)
+        if callable(close_fn):
+            with suppress(Exception):
+                _ = close_fn()
+        self._closed = True
+
+
+def validate_runtime_inputs(source: str, checkpoint: str, config: str) -> None:
+    if source.startswith("cvmmap://") or source.isdigit():
+        pass
+    else:
+        source_path = Path(source)
+        if not source_path.is_file():
+            raise ValueError(f"Video source not found: {source}")
+
+    checkpoint_path = Path(checkpoint)
+    if not checkpoint_path.is_file():
+        raise ValueError(f"Checkpoint not found: {checkpoint}")
+
+    config_path = Path(config)
+    if not config_path.is_file():
+        raise ValueError(f"Config not found: {config}")
+
+
+@click.command(context_settings={"help_option_names": ["-h", "--help"]})
+@click.option("--source", type=str, required=True)
+@click.option("--checkpoint", type=str, required=True)
+@click.option(
+    "--config",
+    type=str,
+    default="configs/sconet/sconet_scoliosis1k.yaml",
+    show_default=True,
+)
+@click.option("--device", type=str, default="cuda:0", show_default=True)
+@click.option("--yolo-model", type=str, default="yolo11n-seg.pt", show_default=True)
+@click.option("--window", type=click.IntRange(min=1), default=30, show_default=True)
+@click.option("--stride", type=click.IntRange(min=1), default=30, show_default=True)
+@click.option("--nats-url", type=str, default=None)
+@click.option(
+    "--nats-subject",
+    type=str,
+    default="scoliosis.result",
+    show_default=True,
+)
+@click.option("--max-frames", type=click.IntRange(min=1), default=None)
+def main(
+    source: str,
+    checkpoint: str,
+    config: str,
+    device: str,
+    yolo_model: str,
+    window: int,
+    stride: int,
+    nats_url: str | None,
+    nats_subject: str,
+    max_frames: int | None,
+) -> None:
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+    )
+
+    try:
+        validate_runtime_inputs(source=source, checkpoint=checkpoint, config=config)
+        pipeline = ScoliosisPipeline(
+            source=source,
+            checkpoint=checkpoint,
+            config=config,
+            device=device,
+            yolo_model=yolo_model,
+            window=window,
+            stride=stride,
+            nats_url=nats_url,
+            nats_subject=nats_subject,
+            max_frames=max_frames,
+        )
+        raise SystemExit(pipeline.run())
+    except ValueError as err:
+        click.echo(f"Error: {err}", err=True)
+        raise SystemExit(2) from err
+    except RuntimeError as err:
+        click.echo(f"Runtime error: {err}", err=True)
+        raise SystemExit(1) from err
diff --git a/opengait/demo/preprocess.py b/opengait/demo/preprocess.py
new file mode 100644
index 0000000..74b9df9
--- /dev/null
+++ b/opengait/demo/preprocess.py
@@ -0,0 +1,270 @@
+from collections.abc import Callable
+import math
+from typing import cast
+
+import cv2
+from beartype import beartype
+import jaxtyping
+from jaxtyping import Float, UInt8
+import numpy as np
+from numpy import ndarray
+from numpy.typing import NDArray
+
+SIL_HEIGHT = 64
+SIL_WIDTH = 44
+SIL_FULL_WIDTH = 64
+SIDE_CUT = 10
+MIN_MASK_AREA = 500
+
+JaxtypedDecorator = Callable[[Callable[..., object]], Callable[..., object]]
+JaxtypedFactory = Callable[..., JaxtypedDecorator]
+jaxtyped = cast(JaxtypedFactory, jaxtyping.jaxtyped)
+
+UInt8Array = NDArray[np.uint8]
+Float32Array = NDArray[np.float32]
+
+
+def _read_attr(container: object, key: str) -> object | None:
+    if isinstance(container, dict):
+        dict_obj = cast(dict[object, object], container)
+        return dict_obj.get(key)
+    try:
+        return cast(object, object.__getattribute__(container, key))
+    except AttributeError:
+        return None
+
+
+def _to_numpy_array(value: object) -> NDArray[np.generic]:
+    current: object = value
+    if isinstance(current, np.ndarray):
+        return current
+
+    detach_obj = _read_attr(current, "detach")
+    if callable(detach_obj):
+        detach_fn = cast(Callable[[], object], detach_obj)
+        current = detach_fn()
+
+    cpu_obj = _read_attr(current, "cpu")
+    if callable(cpu_obj):
+        cpu_fn = cast(Callable[[], object], cpu_obj)
+        current = cpu_fn()
+
+    numpy_obj = _read_attr(current, "numpy")
+    if callable(numpy_obj):
+        numpy_fn = cast(Callable[[], object], numpy_obj)
+        as_numpy = numpy_fn()
+        if isinstance(as_numpy, np.ndarray):
+            return as_numpy
+
+    return cast(NDArray[np.generic], np.asarray(current))
+
+
+def _bbox_from_mask(mask: UInt8[ndarray, "h w"]) -> tuple[int, int, int, int] | None:
+    mask_u8 = np.asarray(mask, dtype=np.uint8)
+    coords = np.argwhere(mask_u8 > 0)
+    if int(coords.size) == 0:
+        return None
+
+    ys = coords[:, 0].astype(np.int64)
+    xs = coords[:, 1].astype(np.int64)
+    x1 = int(np.min(xs))
+    x2 = int(np.max(xs)) + 1
+    y1 = int(np.min(ys))
+    y2 = int(np.max(ys)) + 1
+    if x2 <= x1 or y2 <= y1:
+        return None
+    return (x1, y1, x2, y2)
+
+
+def _sanitize_bbox(
+    bbox: tuple[int, int, int, int], height: int, width: int
+) -> tuple[int, int, int, int] | None:
+    x1, y1, x2, y2 = bbox
+    x1c = max(0, min(int(x1), width - 1))
+    y1c = max(0, min(int(y1), height - 1))
+    x2c = max(0, min(int(x2), width))
+    y2c = max(0, min(int(y2), height))
+    if x2c <= x1c or y2c <= y1c:
+        return None
+    return (x1c, y1c, x2c, y2c)
+
+
+@jaxtyped(typechecker=beartype)
+def frame_to_person_mask(
+    result: object, min_area: int = MIN_MASK_AREA
+) -> tuple[UInt8[ndarray, "h w"], tuple[int, int, int, int]] | None:
+    masks_obj = _read_attr(result, "masks")
+    if masks_obj is None:
+        return None
+
+    masks_data_obj = _read_attr(masks_obj, "data")
+    if masks_data_obj is None:
+        return None
+
+    masks_raw = _to_numpy_array(masks_data_obj)
+    masks_float = np.asarray(masks_raw, dtype=np.float32)
+    if masks_float.ndim == 2:
+        masks_float = masks_float[np.newaxis, ...]
+    if masks_float.ndim != 3:
+        return None
+    mask_count = int(cast(tuple[int, int, int], masks_float.shape)[0])
+    if mask_count <= 0:
+        return None
+
+    box_values: list[tuple[float, float, float, float]] | None = None
+    boxes_obj = _read_attr(result, "boxes")
+    if boxes_obj is not None:
+        xyxy_obj = _read_attr(boxes_obj, "xyxy")
+        if xyxy_obj is not None:
+            xyxy_raw = np.asarray(_to_numpy_array(xyxy_obj), dtype=np.float32)
+            if xyxy_raw.ndim == 1 and int(xyxy_raw.size) >= 4:
+                xyxy_2d = np.asarray(xyxy_raw[:4].reshape(1, 4), dtype=np.float64)
+                x1f = cast(np.float64, xyxy_2d[0, 0])
+                y1f = cast(np.float64, xyxy_2d[0, 1])
+                x2f = cast(np.float64, xyxy_2d[0, 2])
+                y2f = cast(np.float64, xyxy_2d[0, 3])
+                box_values = [
+                    (
+                        float(x1f),
+                        float(y1f),
+                        float(x2f),
+                        float(y2f),
+                    )
+                ]
+            elif xyxy_raw.ndim == 2:
+                shape_2d = cast(tuple[int, int], xyxy_raw.shape)
+                if int(shape_2d[1]) >= 4:
+                    xyxy_2d = np.asarray(xyxy_raw[:, :4], dtype=np.float64)
+                    box_values = []
+                    for row_idx in range(int(cast(tuple[int, int], xyxy_2d.shape)[0])):
+                        x1f = cast(np.float64, xyxy_2d[row_idx, 0])
+                        y1f = cast(np.float64, xyxy_2d[row_idx, 1])
+                        x2f = cast(np.float64, xyxy_2d[row_idx, 2])
+                        y2f = cast(np.float64, xyxy_2d[row_idx, 3])
+                        box_values.append(
+                            (
+                                float(x1f),
+                                float(y1f),
+                                float(x2f),
+                                float(y2f),
+                            )
+                        )
+
+    best_area = -1
+    best_mask: UInt8[ndarray, "h w"] | None = None
+    best_bbox: tuple[int, int, int, int] | None = None
+
+    for idx in range(mask_count):
+        mask_float = np.asarray(masks_float[idx], dtype=np.float32)
+        if mask_float.ndim != 2:
+            continue
+        mask_binary = np.where(mask_float > 0.5, np.uint8(255), np.uint8(0)).astype(
+            np.uint8
+        )
+        mask_u8 = cast(UInt8[ndarray, "h w"], mask_binary)
+
+        area = int(np.count_nonzero(mask_u8))
+        if area < min_area:
+            continue
+
+        bbox: tuple[int, int, int, int] | None = None
+        shape_2d = cast(tuple[int, int], mask_binary.shape)
+        h = int(shape_2d[0])
+        w = int(shape_2d[1])
+        if box_values is not None:
+            box_count = len(box_values)
+            if idx >= box_count:
+                continue
+            row0, row1, row2, row3 = box_values[idx]
+            bbox_candidate = (
+                int(math.floor(row0)),
+                int(math.floor(row1)),
+                int(math.ceil(row2)),
+                int(math.ceil(row3)),
+            )
+            bbox = _sanitize_bbox(bbox_candidate, h, w)
+
+        if bbox is None:
+            bbox = _bbox_from_mask(mask_u8)
+
+        if bbox is None:
+            continue
+
+        if area > best_area:
+            best_area = area
+            best_mask = mask_u8
+            best_bbox = bbox
+
+    if best_mask is None or best_bbox is None:
+        return None
+
+    return best_mask, best_bbox
+
+
+@jaxtyped(typechecker=beartype)
+def mask_to_silhouette(
+    mask: UInt8[ndarray, "h w"],
+    bbox: tuple[int, int, int, int],
+) -> Float[ndarray, "64 44"] | None:
+    mask_u8 = np.where(mask > 0, np.uint8(255), np.uint8(0)).astype(np.uint8)
+    if int(np.count_nonzero(mask_u8)) < MIN_MASK_AREA:
+        return None
+
+    mask_shape = cast(tuple[int, int], mask_u8.shape)
+    h = int(mask_shape[0])
+    w = int(mask_shape[1])
+    bbox_sanitized = _sanitize_bbox(bbox, h, w)
+    if bbox_sanitized is None:
+        return None
+
+    x1, y1, x2, y2 = bbox_sanitized
+    cropped = mask_u8[y1:y2, x1:x2]
+    if cropped.size == 0:
+        return None
+
+    cropped_u8 = np.asarray(cropped, dtype=np.uint8)
+    row_sums = np.sum(cropped_u8, axis=1, dtype=np.int64)
+    row_nonzero = np.nonzero(row_sums > 0)[0].astype(np.int64)
+    if int(row_nonzero.size) == 0:
+        return None
+    top = int(cast(np.int64, row_nonzero[0]))
+    bottom = int(cast(np.int64, row_nonzero[-1])) + 1
+    tightened = cropped[top:bottom, :]
+    if tightened.size == 0:
+        return None
+
+    tight_shape = cast(tuple[int, int], tightened.shape)
+    tight_h = int(tight_shape[0])
+    tight_w = int(tight_shape[1])
+    if tight_h <= 0 or tight_w <= 0:
+        return None
+
+    resized_w = max(1, int(SIL_HEIGHT * (tight_w / tight_h)))
+    resized = np.asarray(
+        cv2.resize(tightened, (resized_w, SIL_HEIGHT), interpolation=cv2.INTER_CUBIC),
+        dtype=np.uint8,
+    )
+
+    if resized_w >= SIL_FULL_WIDTH:
+        start = (resized_w - SIL_FULL_WIDTH) // 2
+        normalized_64 = resized[:, start : start + SIL_FULL_WIDTH]
+    else:
+        pad_left = (SIL_FULL_WIDTH - resized_w) // 2
+        pad_right = SIL_FULL_WIDTH - resized_w - pad_left
+        normalized_64 = np.pad(
+            resized,
+            ((0, 0), (pad_left, pad_right)),
+            mode="constant",
+            constant_values=0,
+        )
+
+    silhouette = np.asarray(
+        normalized_64[:, SIDE_CUT : SIL_FULL_WIDTH - SIDE_CUT], dtype=np.float32
+    )
+    if silhouette.shape != (SIL_HEIGHT, SIL_WIDTH):
+        return None
+
+    silhouette_norm = np.clip(silhouette / np.float32(255.0), 0.0, 1.0).astype(
+        np.float32
+    )
+    return cast(Float[ndarray, "64 44"], silhouette_norm)
diff --git a/opengait/demo/sconet_demo.py b/opengait/demo/sconet_demo.py
new file mode 100644
index 0000000..377649c
--- /dev/null
+++ b/opengait/demo/sconet_demo.py
@@ -0,0 +1,317 @@
+from __future__ import annotations
+
+from collections.abc import Callable
+from pathlib import Path
+import sys
+from typing import ClassVar, Protocol, cast, override
+
+import torch
+import torch.nn as nn
+from beartype import beartype
+from einops import rearrange
+from jaxtyping import Float
+import jaxtyping
+from torch import Tensor
+
+_OPENGAIT_PACKAGE_ROOT = Path(__file__).resolve().parents[1]
+if str(_OPENGAIT_PACKAGE_ROOT) not in sys.path:
+    sys.path.insert(0, str(_OPENGAIT_PACKAGE_ROOT))
+
+from opengait.modeling.backbones.resnet import ResNet9
+from opengait.modeling.modules import (
+    HorizontalPoolingPyramid,
+    PackSequenceWrapper as TemporalPool,
+    SeparateBNNecks,
+    SeparateFCs,
+)
+from opengait.utils import common as common_utils
+
+
+JaxtypedDecorator = Callable[[Callable[..., object]], Callable[..., object]]
+JaxtypedFactory = Callable[..., JaxtypedDecorator]
+jaxtyped = cast(JaxtypedFactory, jaxtyping.jaxtyped)
+ConfigLoader = Callable[[str], dict[str, object]]
+config_loader = cast(ConfigLoader, common_utils.config_loader)
+
+
+class TemporalPoolLike(Protocol):
+    def __call__(
+        self,
+        seqs: Tensor,
+        seqL: object,
+        dim: int = 2,
+        options: dict[str, int] | None = None,
+    ) -> object: ...
+
+
+class HppLike(Protocol):
+    def __call__(self, x: Tensor) -> Tensor: ...
+
+
+class FCsLike(Protocol):
+    def __call__(self, x: Tensor) -> Tensor: ...
+
+
+class BNNecksLike(Protocol):
+    def __call__(self, x: Tensor) -> tuple[Tensor, Tensor]: ...
+
+
+class ScoNetDemo(nn.Module):
+    LABEL_MAP: ClassVar[dict[int, str]] = {0: "negative", 1: "neutral", 2: "positive"}
+    cfg_path: str
+    cfg: dict[str, object]
+    backbone: ResNet9
+    temporal_pool: TemporalPoolLike
+    hpp: HppLike
+    fcs: FCsLike
+    bn_necks: BNNecksLike
+    device: torch.device
+
+    @jaxtyped(typechecker=beartype)
+    def __init__(
+        self,
+        cfg_path: str | Path = "configs/sconet/sconet_scoliosis1k.yaml",
+        checkpoint_path: str | Path | None = None,
+        device: str | torch.device | None = None,
+    ) -> None:
+        super().__init__()
+        resolved_cfg = self._resolve_path(cfg_path)
+        self.cfg_path = str(resolved_cfg)
+
+        self.cfg = config_loader(self.cfg_path)
+
+        model_cfg = self._extract_model_cfg(self.cfg)
+        backbone_cfg = self._extract_dict(model_cfg, "backbone_cfg")
+
+        if backbone_cfg.get("type") != "ResNet9":
+            raise ValueError(
+                "ScoNetDemo currently supports backbone type ResNet9 only."
+            )
+
+        self.backbone = ResNet9(
+            block=self._extract_str(backbone_cfg, "block"),
+            channels=self._extract_int_list(backbone_cfg, "channels"),
+            in_channel=self._extract_int(backbone_cfg, "in_channel", default=1),
+            layers=self._extract_int_list(backbone_cfg, "layers"),
+            strides=self._extract_int_list(backbone_cfg, "strides"),
+            maxpool=self._extract_bool(backbone_cfg, "maxpool", default=True),
+        )
+
+        fcs_cfg = self._extract_dict(model_cfg, "SeparateFCs")
+        bn_cfg = self._extract_dict(model_cfg, "SeparateBNNecks")
+        bin_num = self._extract_int_list(model_cfg, "bin_num")
+
+        self.temporal_pool = cast(TemporalPoolLike, TemporalPool(torch.max))
+        self.hpp = cast(HppLike, HorizontalPoolingPyramid(bin_num=bin_num))
+        self.fcs = cast(
+            FCsLike,
+            SeparateFCs(
+                parts_num=self._extract_int(fcs_cfg, "parts_num"),
+                in_channels=self._extract_int(fcs_cfg, "in_channels"),
+                out_channels=self._extract_int(fcs_cfg, "out_channels"),
+                norm=self._extract_bool(fcs_cfg, "norm", default=False),
+            ),
+        )
+        self.bn_necks = cast(
+            BNNecksLike,
+            SeparateBNNecks(
+                parts_num=self._extract_int(bn_cfg, "parts_num"),
+                in_channels=self._extract_int(bn_cfg, "in_channels"),
+                class_num=self._extract_int(bn_cfg, "class_num"),
+                norm=self._extract_bool(bn_cfg, "norm", default=True),
+                parallel_BN1d=self._extract_bool(bn_cfg, "parallel_BN1d", default=True),
+            ),
+        )
+
+        self.device = (
+            torch.device(device) if device is not None else torch.device("cpu")
+        )
+        _ = self.to(self.device)
+
+        if checkpoint_path is not None:
+            _ = self.load_checkpoint(checkpoint_path)
+
+        _ = self.eval()
+
+    @staticmethod
+    def _resolve_path(path: str | Path) -> Path:
+        candidate = Path(path)
+        if candidate.is_file():
+            return candidate
+        if candidate.is_absolute():
+            return candidate
+        repo_root = Path(__file__).resolve().parents[2]
+        return repo_root / candidate
+
+    @staticmethod
+    def _extract_model_cfg(cfg: dict[str, object]) -> dict[str, object]:
+        model_cfg_obj = cfg.get("model_cfg")
+        if not isinstance(model_cfg_obj, dict):
+            raise TypeError("model_cfg must be a dictionary.")
+        return cast(dict[str, object], model_cfg_obj)
+
+    @staticmethod
+    def _extract_dict(container: dict[str, object], key: str) -> dict[str, object]:
+        value = container.get(key)
+        if not isinstance(value, dict):
+            raise TypeError(f"{key} must be a dictionary.")
+        return cast(dict[str, object], value)
+
+    @staticmethod
+    def _extract_str(container: dict[str, object], key: str) -> str:
+        value = container.get(key)
+        if not isinstance(value, str):
+            raise TypeError(f"{key} must be a string.")
+        return value
+
+    @staticmethod
+    def _extract_int(
+        container: dict[str, object], key: str, default: int | None = None
+    ) -> int:
+        value = container.get(key, default)
+        if not isinstance(value, int):
+            raise TypeError(f"{key} must be an int.")
+        return value
+
+    @staticmethod
+    def _extract_bool(
+        container: dict[str, object], key: str, default: bool | None = None
+    ) -> bool:
+        value = container.get(key, default)
+        if not isinstance(value, bool):
+            raise TypeError(f"{key} must be a bool.")
+        return value
+
+    @staticmethod
+    def _extract_int_list(container: dict[str, object], key: str) -> list[int]:
+        value = container.get(key)
+        if not isinstance(value, list):
+            raise TypeError(f"{key} must be a list[int].")
+        values = cast(list[object], value)
+        if not all(isinstance(v, int) for v in values):
+            raise TypeError(f"{key} must be a list[int].")
+        return cast(list[int], values)
+
+    @staticmethod
+    def _normalize_state_dict(
+        state_dict_obj: dict[object, object],
+    ) -> dict[str, Tensor]:
+        prefix_remap: tuple[tuple[str, str], ...] = (
+            ("Backbone.forward_block.", "backbone."),
+            ("FCs.", "fcs."),
+            ("BNNecks.", "bn_necks."),
+        )
+        cleaned_state_dict: dict[str, Tensor] = {}
+        for key_obj, value_obj in state_dict_obj.items():
+            if not isinstance(key_obj, str):
+                raise TypeError("Checkpoint state_dict keys must be strings.")
+            if not isinstance(value_obj, Tensor):
+                raise TypeError("Checkpoint state_dict values must be torch.Tensor.")
+            key = key_obj[7:] if key_obj.startswith("module.") else key_obj
+            for source_prefix, target_prefix in prefix_remap:
+                if key.startswith(source_prefix):
+                    key = f"{target_prefix}{key[len(source_prefix) :]}"
+                    break
+            if key in cleaned_state_dict:
+                raise RuntimeError(
+                    f"Checkpoint key normalization collision detected for key '{key}'."
+                )
+            cleaned_state_dict[key] = value_obj
+        return cleaned_state_dict
+
+    @jaxtyped(typechecker=beartype)
+    def load_checkpoint(
+        self,
+        checkpoint_path: str | Path,
+        map_location: str | torch.device | None = None,
+        strict: bool = True,
+    ) -> None:
+        resolved_ckpt = self._resolve_path(checkpoint_path)
+        checkpoint_obj = cast(
+            object,
+            torch.load(
+                str(resolved_ckpt),
+                map_location=map_location if map_location is not None else self.device,
+            ),
+        )
+
+        state_dict_obj: object = checkpoint_obj
+        if isinstance(checkpoint_obj, dict) and "model" in checkpoint_obj:
+            state_dict_obj = cast(dict[str, object], checkpoint_obj)["model"]
+
+        if not isinstance(state_dict_obj, dict):
+            raise TypeError("Unsupported checkpoint format.")
+
+        cleaned_state_dict = self._normalize_state_dict(
+            cast(dict[object, object], state_dict_obj)
+        )
+        try:
+            _ = self.load_state_dict(cleaned_state_dict, strict=strict)
+        except RuntimeError as exc:
+            raise RuntimeError(
+                f"Failed to load ScoNetDemo checkpoint after key normalization from '{resolved_ckpt}'."
+            ) from exc
+        _ = self.eval()
+
+    def _prepare_sils(self, sils: Tensor) -> Tensor:
+        if sils.ndim == 4:
+            sils = sils.unsqueeze(1)
+        elif sils.ndim == 5 and sils.shape[1] != 1 and sils.shape[2] == 1:
+            sils = rearrange(sils, "b s c h w -> b c s h w")
+
+        if sils.ndim != 5 or sils.shape[1] != 1:
+            raise ValueError("Expected sils shape [B, 1, S, H, W] or [B, S, H, W].")
+
+        return sils.float().to(self.device)
+
+    def _forward_backbone(self, sils: Tensor) -> Tensor:
+        batch, channels, seq, height, width = sils.shape
+        framewise = sils.transpose(1, 2).reshape(batch * seq, channels, height, width)
+        frame_feats = cast(Tensor, self.backbone(framewise))
+        _, out_channels, out_h, out_w = frame_feats.shape
+        return (
+            frame_feats.reshape(batch, seq, out_channels, out_h, out_w)
+            .transpose(1, 2)
+            .contiguous()
+        )
+
+    @override
+    @jaxtyped(typechecker=beartype)
+    def forward(self, sils: Float[Tensor, "batch 1 seq 64 44"]) -> dict[str, Tensor]:
+        with torch.inference_mode():
+            prepared_sils = self._prepare_sils(sils)
+            outs = self._forward_backbone(prepared_sils)
+
+            pooled_obj = self.temporal_pool(outs, None, options={"dim": 2})
+            if (
+                not isinstance(pooled_obj, tuple)
+                or not pooled_obj
+                or not isinstance(pooled_obj[0], Tensor)
+            ):
+                raise TypeError("TemporalPool output is invalid.")
+            pooled = pooled_obj[0]
+
+            feat = self.hpp(pooled)
+            embed_1 = self.fcs(feat)
+            _, logits = self.bn_necks(embed_1)
+
+            mean_logits = logits.mean(dim=-1)
+            pred_ids = torch.argmax(mean_logits, dim=-1)
+            probs = torch.softmax(mean_logits, dim=-1)
+            confidence = torch.gather(
+                probs, dim=-1, index=pred_ids.unsqueeze(-1)
+            ).squeeze(-1)
+
+            return {"logits": logits, "label": pred_ids, "confidence": confidence}
+
+    @jaxtyped(typechecker=beartype)
+    def predict(self, sils: Float[Tensor, "batch 1 seq 64 44"]) -> tuple[str, float]:
+        outputs = cast(dict[str, Tensor], self.forward(sils))
+        labels = outputs["label"]
+        confidence = outputs["confidence"]
+
+        if labels.numel() != 1:
+            raise ValueError("predict expects batch size 1.")
+
+        label_id = int(labels.item())
+        return self.LABEL_MAP[label_id], float(confidence.item())
diff --git a/opengait/demo/window.py b/opengait/demo/window.py
new file mode 100644
index 0000000..c93eda3
--- /dev/null
+++ b/opengait/demo/window.py
@@ -0,0 +1,295 @@
+"""Sliding window / ring buffer manager for real-time gait analysis.
+
+This module provides bounded buffer management for silhouette sequences
+with track ID tracking and gap detection.
+"""
+
+from collections import deque
+from typing import TYPE_CHECKING, Protocol, final
+
+import numpy as np
+import torch
+from jaxtyping import Float
+from numpy import ndarray
+
+if TYPE_CHECKING:
+    from numpy.typing import NDArray
+
+
+# Silhouette dimensions from preprocess.py
+SIL_HEIGHT: int = 64
+SIL_WIDTH: int = 44
+
+
+class _Boxes(Protocol):
+    """Protocol for boxes with xyxy and id attributes."""
+
+    @property
+    def xyxy(self) -> "NDArray[np.float32] | object": ...
+    @property
+    def id(self) -> "NDArray[np.int64] | object | None": ...
+
+
+class _Masks(Protocol):
+    """Protocol for masks with data attribute."""
+
+    @property
+    def data(self) -> "NDArray[np.float32] | object": ...
+
+
+class _DetectionResults(Protocol):
+    """Protocol for detection results from Ultralytics-style objects."""
+
+    @property
+    def boxes(self) -> _Boxes: ...
+    @property
+    def masks(self) -> _Masks: ...
+
+
+@final
+class SilhouetteWindow:
+    """Bounded sliding window for silhouette sequences.
+
+    Manages a fixed-size buffer of silhouettes with track ID tracking
+    and automatic reset on track changes or frame gaps.
+
+    Attributes:
+        window_size: Maximum number of frames in the buffer.
+        stride: Classification stride (frames between classifications).
+        gap_threshold: Maximum allowed frame gap before reset.
+    """
+
+    window_size: int
+    stride: int
+    gap_threshold: int
+    _buffer: deque[Float[ndarray, "64 44"]]
+    _frame_indices: deque[int]
+    _track_id: int | None
+    _last_classified_frame: int
+    _frame_count: int
+
+    def __init__(
+        self,
+        window_size: int = 30,
+        stride: int = 1,
+        gap_threshold: int = 15,
+    ) -> None:
+        """Initialize the silhouette window.
+
+        Args:
+            window_size: Maximum buffer size (default 30).
+            stride: Frames between classifications (default 1).
+            gap_threshold: Max frame gap before reset (default 15).
+        """
+        self.window_size = window_size
+        self.stride = stride
+        self.gap_threshold = gap_threshold
+
+        # Bounded storage via deque
+        self._buffer = deque(maxlen=window_size)
+        self._frame_indices = deque(maxlen=window_size)
+        self._track_id = None
+        self._last_classified_frame = -1
+        self._frame_count = 0
+
+    def push(self, sil: np.ndarray, frame_idx: int, track_id: int) -> None:
+        """Push a new silhouette into the window.
+
+        Automatically resets buffer on track ID change or frame gap
+        exceeding gap_threshold.
+
+        Args:
+            sil: Silhouette array of shape (64, 44), float32.
+            frame_idx: Current frame index for gap detection.
+            track_id: Track ID for the person.
+        """
+        # Check for track ID change
+        if self._track_id is not None and track_id != self._track_id:
+            self.reset()
+
+        # Check for frame gap
+        if self._frame_indices:
+            last_frame = self._frame_indices[-1]
+            gap = frame_idx - last_frame
+            if gap > self.gap_threshold:
+                self.reset()
+
+        # Update track ID
+        self._track_id = track_id
+
+        # Validate and append silhouette
+        sil_array = np.asarray(sil, dtype=np.float32)
+        if sil_array.shape != (SIL_HEIGHT, SIL_WIDTH):
+            raise ValueError(
+                f"Expected silhouette shape ({SIL_HEIGHT}, {SIL_WIDTH}), got {sil_array.shape}"
+            )
+
+        self._buffer.append(sil_array)
+        self._frame_indices.append(frame_idx)
+        self._frame_count += 1
+
+    def is_ready(self) -> bool:
+        """Check if window has enough frames for classification.
+
+        Returns:
+            True if buffer is full (window_size frames).
+        """
+        return len(self._buffer) >= self.window_size
+
+    def should_classify(self) -> bool:
+        """Check if classification should run based on stride.
+
+        Returns:
+            True if enough frames have passed since last classification.
+        """
+        if not self.is_ready():
+            return False
+
+        if self._last_classified_frame < 0:
+            return True
+
+        current_frame = self._frame_indices[-1]
+        frames_since = current_frame - self._last_classified_frame
+        return frames_since >= self.stride
+
+    def get_tensor(self, device: str = "cpu") -> torch.Tensor:
+        """Get window contents as a tensor for model input.
+
+        Args:
+            device: Target device for the tensor (default 'cpu').
+
+        Returns:
+            Tensor of shape [1, 1, window_size, 64, 44] with dtype float32.
+
+        Raises:
+            ValueError: If buffer is not full.
+        """
+        if not self.is_ready():
+            raise ValueError(
+                f"Window not ready: {len(self._buffer)}/{self.window_size} frames"
+            )
+
+        # Stack buffer into array [window_size, 64, 44]
+        stacked = np.stack(list(self._buffer), axis=0)
+
+        # Add batch and channel dims: [1, 1, window_size, 64, 44]
+        tensor = torch.from_numpy(stacked.astype(np.float32))
+        tensor = tensor.unsqueeze(0).unsqueeze(0)
+
+        return tensor.to(device)
+
+    def reset(self) -> None:
+        """Reset the window, clearing all buffers and counters."""
+        self._buffer.clear()
+        self._frame_indices.clear()
+        self._track_id = None
+        self._last_classified_frame = -1
+        self._frame_count = 0
+
+    def mark_classified(self) -> None:
+        """Mark current frame as classified, updating stride tracking."""
+        if self._frame_indices:
+            self._last_classified_frame = self._frame_indices[-1]
+
+    @property
+    def current_track_id(self) -> int | None:
+        """Current track ID, or None if buffer is empty."""
+        return self._track_id
+
+    @property
+    def frame_count(self) -> int:
+        """Total frames pushed since last reset."""
+        return self._frame_count
+
+    @property
+    def fill_level(self) -> float:
+        """Fill ratio of the buffer (0.0 to 1.0)."""
+        return len(self._buffer) / self.window_size
+
+
+def select_person(
+    results: _DetectionResults,
+) -> tuple[ndarray, tuple[int, int, int, int], int] | None:
+    """Select the person with largest bounding box from detection results.
+
+    Args:
+        results: Detection results object with boxes and masks attributes.
+            Expected to have:
+            - boxes.xyxy: array of bounding boxes [N, 4]
+            - masks.data: array of masks [N, H, W]
+            - boxes.id: optional track IDs [N]
+
+    Returns:
+        Tuple of (mask, bbox, track_id) for the largest person,
+        or None if no valid detections or track IDs unavailable.
+    """
+    # Check for track IDs
+    boxes_obj: _Boxes | object = getattr(results, "boxes", None)
+    if boxes_obj is None:
+        return None
+
+    track_ids_obj: ndarray | object | None = getattr(boxes_obj, "id", None)
+    if track_ids_obj is None:
+        return None
+
+    track_ids: ndarray = np.asarray(track_ids_obj)
+    if track_ids.size == 0:
+        return None
+
+    # Get bounding boxes
+    xyxy_obj: ndarray | object = getattr(boxes_obj, "xyxy", None)
+    if xyxy_obj is None:
+        return None
+
+    bboxes: ndarray = np.asarray(xyxy_obj)
+    if bboxes.ndim == 1:
+        bboxes = bboxes.reshape(1, -1)
+
+    if bboxes.shape[0] == 0:
+        return None
+
+    # Get masks
+    masks_obj: _Masks | object = getattr(results, "masks", None)
+    if masks_obj is None:
+        return None
+
+    masks_data: ndarray | object = getattr(masks_obj, "data", None)
+    if masks_data is None:
+        return None
+
+    masks: ndarray = np.asarray(masks_data)
+    if masks.ndim == 2:
+        masks = masks[np.newaxis, ...]
+
+    if masks.shape[0] != bboxes.shape[0]:
+        return None
+
+    # Find largest bbox by area
+    best_idx: int = -1
+    best_area: float = -1.0
+
+    for i in range(int(bboxes.shape[0])):
+        row: "NDArray[np.float32]" = bboxes[i][:4]
+        x1f: float = float(row[0])
+        y1f: float = float(row[1])
+        x2f: float = float(row[2])
+        y2f: float = float(row[3])
+        area: float = (x2f - x1f) * (y2f - y1f)
+        if area > best_area:
+            best_area = area
+            best_idx = i
+
+    if best_idx < 0:
+        return None
+
+    # Extract mask and bbox
+    mask: "NDArray[np.float32]" = masks[best_idx]
+    bbox = (
+        int(float(bboxes[best_idx][0])),
+        int(float(bboxes[best_idx][1])),
+        int(float(bboxes[best_idx][2])),
+        int(float(bboxes[best_idx][3])),
+    )
+    track_id = int(track_ids[best_idx]) if best_idx < len(track_ids) else best_idx
+
+    return mask, bbox, track_id