Monitoring & cameras

• Multiple “cameras” — each camera is a watched folder with its own name, paths, and behavior.
• Automatic processing — new images in a watch folder trigger object detection without manual steps.
• Fine-grained camera controls — enable/disable cameras, optional filename prefix filters, and per-camera caps on how many inferences run per minute.
• Skip stale files — ignore images that are too old so bursts or replays do not flood the pipeline.
• Timed pauses — pause processing on selected cameras for a set duration (e.g. maintenance or known activity).
• Detection masks — restrict where in the image objects are counted (ignore sky, roads, UI chrome, etc.).
• Time-of-day detection profiles — different class filters, confidence, max objects, and half-precision settings by time band over the day.

AI models

• YOLO-based detection — run modern YOLO-family models for fast, accurate object detection.
• PyTorch and ONNX — use .pt models or .onnx for flexible deployment and hardware options.
• Model schedule — assign cameras to groups and switch which model is active by time of day (e.g. day vs night).
• Model manager in the browser — upload models, convert .pt → ONNX, browse Hugging Face for compatible weights, and download from the UI.
• Try before production — test inference on a single image (e.g. drag-and-drop) to validate models and settings.

Results & visibility

• Saved outputs — annotated images plus structured JSON results per event; optional layout with a global output area and per-camera subfolders.
• Detections gallery — browse history with previews, full result details, and filter by camera.
• Statistics — per-camera performance and timing trends (charts) to see how the system is behaving.
• Logs — operational and processing logs with live streaming and filtering in the UI.
• Hardware view — see system/GPU-oriented information relevant to running inference.
• At-a-glance status — dashboard indicator for whether the server is up and watchers are actively processing.

Alerts & automation

• MQTT — publish detection results per camera; send JSON, images, or both; templated messages with placeholders; subscribe to an MQTT image feed so images can be processed from the broker, not only from disk.
• Telegram — send alerts (text and/or photos with captions) to chosen chats per camera.
• Telegram bot (optional) — define custom commands for things like pausing/resuming cameras, querying last detection, hitting HTTP endpoints, publishing MQTT, or running approved shell-style automations (with safety controls).

Security & access

• IP allowlist — restrict who can reach the web UI (e.g. localhost only or your LAN/VPN ranges).

Deployment & operations

• Browser-based configuration — no separate desktop config app; everything is managed through the web UI.
• Cross-platform — runs on Windows, macOS, and Linux.
• Runs in the background — optional Windows service, or macOS launchd / Linux systemd so monitoring survives logouts and reboots.
• Self-contained data — settings, models, and logs live under the app directory for straightforward backup and migration.
• In-app updates (where supported by your build) — streamlined upgrade path without re-explaining licensing here.