Every frame moves through four stages — detection finds the object, tracking holds it, color fusion confirms it, and your hardware acts on it. Every stage is yours to tune from a clean local dashboard, in real time.
Click between real-time vision servoing, dataset labeling with SentinelLabeler, empirical silicon benchmarks, and zero-GC developer APIs.
Built specifically for capture card video streams. SentinelLabeler eliminates tedious manual frame-by-frame labeling with cutting-edge AI assistance:
Click once on any target to instantly generate tight polygon segmentation masks with high edge fidelity.
Annotate a single frame and propagate bounding boxes across up to 300 sequential video frames automatically.
Outputs clean normalized YOLO text labels and train/val directories ready for direct PyTorch / ONNX training.
// ============================================================================
// Sentinel Core Vision — Zero-GC 1,000Hz Hardware Dispatch
// High-Frequency Visual Servoing & Kinematic State Estimation
// ============================================================================
[StructLayout(LayoutKind.Sequential, Pack = 1)]
public readonly record struct DirectHidPacket(
byte ReportId, // 0x01 = Direct Controller Emulation
short StickLeftX, // [-32768, 32767] Normalized Deflection
short StickLeftY,
short StickRightX, // Servoing Saccade Step
short StickRightY,
uint ButtonsMask, // Raw Gamepad Triggers & ADS
uint TimestampMicro // 1,000Hz Jitter Compensation
);
public sealed class ServoKinematicPipeline : IDisposable
{
private readonly OutputLoop _outputLoop;
private readonly ProjectionMagnetCleanAimCore _aimCore;
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public void DispatchServoTick(in ReadOnlySpan targets, in VideoFrameHeader frame)
{
// Zero allocation pass: IMM filter estimates target velocity & acceleration
ref readonly var activeTarget = ref _aimCore.EvaluateBestTarget(targets, frame.Timestamp);
// Time-optimal phase-plane snapping with anti-deadzone compensation
Vector2 stickDeflection = _aimCore.ComputeKinematicStep(activeTarget);
// Zero-GC atomic transmission to Titan Two hardware via USB DirectHID
_outputLoop.EnqueueStickDeflection(stickDeflection.X, stickDeflection.Y);
}
}
A fast ONNX vision model scans every captured frame with GPU acceleration across multiple execution providers — CUDA, TensorRT, DirectML, or OpenVINO — and surfaces the result on a status-first dashboard so you always know the engine is healthy. Because it runs locally, it's designed for low-latency processing, and you read the live system state at a glance instead of guessing.
The Vision page puts the raw feed and the AI overlay together — detections, tracking, ownership, and color, all drawn live as the engine decides. It turns a black box into a window, so when you change a setting you can watch the effect instead of trusting it blind.
A live color scout ranks the colors in your frame so the object outline is obvious, then HSV color confirmation runs alongside the model to agree on the same object. It keeps you locked on the object of interest and off any ignored class — and the sticky preview sits right beside the controls so you tune it by eye.
Start with a handful of plain-language sliders for feel, then reveal advanced and full developer surfaces only when you want them. Nothing lives in a config file — every parameter is a control in the browser, so you can go as deep as you like without ever leaving the dashboard.
Setup walks you through your capture source and output device and checks the whole chain is ready before you start. Output goes through a supported output device so it reads like native input — no drivers injected into your application — which keeps your system clean and your input native.
The usual alternative is a hand-written script and a lot of trial and error. Here's the difference.
The details that make it feel finished — and keep everything on your own machine.
Updated September 24, 2026
Kinetic, Glide, Magnet, and Obsidian — each with its own color identity in the dashboard and its own tuning layout. Pick the one that matches how you play.
Try a candidate setting, keep it if it feels better, or roll it back in one click. Export a support bundle if something still looks off.
Prepare and flash GPC scripts, read slot state, and watch live Titan telemetry — no separate Titan software required.
Run Sentinel through a virtual Xbox controller (ViGEm) or directly through an existing Xbox Remote Play session — no extra hardware required.
Jittery, overshooting, grabbing the wrong target, too sticky, dead zone too tight — pick the symptom and the dashboard points you to the right control, plus what to check next.
Load a different ONNX model on the fly — no restart, no downtime. Compare detection models in seconds.
Capture an entire tuned setup as a profile and switch between them per source or per feel in one click.
Choose how tracking behaves — from a bare baseline to predictive, belief-following controllers — live.
Define where the engine is allowed to look and act, with adjustable field-of-view and engagement zones.
Works from a real capture feed at the device's native resolution and refresh — no special drivers required.
The engine runs entirely on your PC. The only network call verifies your license — never your activity.
A native Windows companion app launches and manages the engine, with native WinRT capture and an output loop tunable up to 1000Hz.
Activate in one click and dial in your first profile in minutes — every stage, in your browser.