Technology

How Sentinel Core Vision works

A real-time vision pipeline that runs entirely on your machine — every stage visible, swappable, and tunable.

The pipeline

Six stages, one continuous loop

Every frame flows through the same path — from capture to output — and each stage is exposed for inspection and tuning.

1 · Capture

Frames come in from a screen, a camera, or a capture card through native WinRT capture — at the source's native resolution and refresh, with the capture mode auto-selected for your setup.

WinRT · screen · camera · capture card

2 · Preprocess

Each frame is resized and normalized into the shape the model expects — fast, GPU-assisted frame prep that keeps the loop tight.

GPU-assisted frame prep

3 · Inference

An ONNX Runtime model scans the prepared frame and returns detections — the objects it sees and where they are.

ONNX Runtime models

4 · Tracking

Detections are stitched into persistent identities across frames. Choose from multiple swappable tracking engines with predictive motion.

Multiple swappable engines

5 · Color fusion

An optional HSV color check runs alongside the model, confirming the object by its color before anything acts on it.

Optional HSV confirmation

6 · Output

The result is delivered with low latency to a supported input device, so it reads like native input. The output loop's decision rate is fully configurable, up to 1000Hz.

Configurable up to 1000Hz
On-device

Runs on your hardware

The whole pipeline lives on your PC — and it's built to make the most of the GPU you already have.

Local-first

100% on-device processing

Every frame is captured, prepared, analyzed, and acted on right on your machine. Nothing about what you're looking at is uploaded — frames never leave your PC. The vision loop has no cloud dependency; the only outside traffic is your license check and, if you choose to grab a model from the marketplace, downloading that model file from our cloud-hosted catalog.

  • Frames are processed locally, never uploaded
  • No cloud dependency in the vision loop itself
  • Cloud is used only for licensing and the model marketplace
Execution providers

Multiple execution providers

Sentinel runs on ONNX Runtime and picks the right acceleration path for your machine: NVIDIA cards via CUDA and TensorRT, AMD and Intel via DirectML, or OpenVINO. A CPU execution provider is also available as a last-resort fallback, though it isn't recommended for real-time use. When your preferred provider isn't available, the engine falls back automatically so it keeps running.

  • NVIDIA via CUDA / TensorRT
  • AMD & Intel via DirectML, or OpenVINO
  • CPU fallback available (not recommended)
Models

Hot-swappable ONNX models

Load a different ONNX model on the fly — no restart, no downtime. Swap detection models in seconds to compare accuracy and speed, and keep the one that fits your setup best.

  • Drop-in ONNX model loading
  • No restart to change models
  • Compare models side by side quickly
Dashboard

A local dashboard at localhost:5000

The whole engine is driven from a dashboard in your browser at http://localhost:5000. Watch live system health, inspect every decision on the vision view, and tune each stage in real time — no config files, no command line.

  • Live tuning and inspection in the browser
  • System health and pipeline state at a glance
  • Every parameter is a labeled control
Sentinel Core Vision Studio

A native app that runs the whole show

Studio is a native Windows companion app that launches and manages the engine — provider selection, capture setup, and a readiness check happen in one guided window before the dashboard ever opens. It's built on native WinRT capture, so screens, windows, and capture devices are detected and selected the way Windows itself sees them.

  • Native Windows app, not a wrapped browser
  • Guided provider & capture setup on launch
  • WinRT capture for reliable, native-resolution frames
Output loop

Configurable up to 1000Hz

The output stage's decision rate isn't fixed — it's a control you set. Two selectable projection engines drive it: a proven stable baseline, and a newer predictive engine with an optional continuity mode that holds its lead steady through target changes instead of resetting on every switch.

  • Output decision rate tunable up to 1000Hz
  • Two selectable projection engines
  • Optional WorldState motion bridge for smoother handoffs
Performance

Built for low latency

The pipeline is engineered for high-frequency, low-latency operation — from frame in to output out.

The stack it runs on
ONNX Runtime NVIDIA CUDA TensorRT DirectML (AMD & Intel) OpenVINO CPU fallback HSV color fusion Multiple tracking engines WinRT capture 1000Hz output loop Native Studio app

Real-world latency and frame rate depend on your hardware — your GPU, drivers, the model you run, your capture resolution, and your configuration all affect the numbers you see.

See the whole pipeline for yourself

Activate in one click and watch every stage run live in your browser — capture to output, all on your machine.