Live predictions, hold-out signatures
Click any of the 10 hold-out signatures to run a live genuine-vs-forged prediction with our custom neural network in real-time (ML Inference, Azure, Serverless). Hold-out means that these images were set aside and never shown to the network during training, so each click is a real prediction on unseen input.
Custom neural network · built from scratch · POC
Why a small custom network? Fine-tuning a pre-trained LLM (vision model) on this data gets the same result. But, it is overkill with an order of magnitude increase in compute cost for training and inference. This task is narrow: it's a binary classification. A small network built from scratch learns exactly that, trains in minutes, and runs fast on serverless. No GPU bill, no giant model to host. Same accuracy, far simpler to run.
ML Data Preparation: Normalisation and augmentation are applied to the dataset to suppress spurious background and texture cues. This biases the training process toward the geometric form of the signature itself, rather than the incidental properties of the sample images.