On-device inference for computer vision and audio: ONNX runtime, CNN architectures sized for constrained hardware, camera calibration, and the training tricks that keep edge AI models honest outside the lab.
Modern AI products feel personal only when they remember, that bug you were debugging yesterday, the pet profile you set up last month, the long-running coaching thread you keep coming back to. Under the hood, though, LLM inference APIs don't remember anything. Every call is a clean slate,
How AI models are shrinking without losing their minds. Many of our real-time models, from detecting dog motion to analyzing eating patterns, run directly on embedded systems. At Hoomanely, we pair lightweight CNNs like MobileNet with quantization and pruning to get low-latency inference without giving up reliability. Processing data locally
Running inference on-device — not in the cloud — means every model decision is also a hardware decision: memory budget, latency budget, and power budget all push back on accuracy. This hub covers what it takes to get computer vision and audio models running on constrained edge hardware: runtime selection, calibration, depth estimation, and the training tricks that keep a model honest outside the lab.
We build this way because our pet devices need to respond in real time, offline, on hardware that costs a few dollars — a cloud round-trip isn't an option when a dog is already at the bowl. That constraint shapes everything from model architecture to how we validate training data.
ONNX Runtime: The Engine Behind Flexible AI Inference covers the runtime layer that lets one model target multiple chips. Choosing the Right Model Architecture for Production and Continual YOLO Training on the Edge cover the architecture and retraining decisions behind our detection models. For vision specifically, Aligning the Invisible: Stereo Camera Sensor Calibration and Monocular Depth Estimation: Depth From a Single Camera cover how we recover 3D structure from cheap cameras. Dog Identification by Facial Landmarks: Real-World Results shows what that pipeline looks like shipped, and MelCNN for Edge Audio Intelligence covers the audio side of edge inference — the same constraints, a different signal.
If you're trying to make AI work on hardware that can't afford a GPU, this is the hub for you.