Insidious AI — Real-Time Vision Desktop App

System design

 Screen capture (frame loop)
        │
        ▼
 Pre-process (crop · resize · normalise)
        │
        ▼
 ONNX model  ──► DirectML (GPU)  /  CPU fallback
        │
        ▼
 Post-process (boxes · confidence · target selection)
        │
        ▼
 Input layer ── mouse  |  PS5 DualSense / Xbox controller
        ▲
 WPF UI: settings, model selection, overlays

What I changed

• Complete UI redesign of the WPF application
• CPU-first model initialisation for reliable start-up on machines without a compatible GPU
• Native controller support (PS5 DualSense and Xbox), not just mouse input
• Built on the DirectML fork of Aimmy for faster inference; credit to the original authors

Engineering takeaway

Real-time inference is a latency budget: every stage of capture → model → input has to fit inside a frame.