Vision Studio

Squint Vision Studio maps a trained model's embedding space, marks the regions where it can be trusted and not trusted, and exports that map so every decision is precise.

Request a Demo

Read the Documentation

The Squint Vision Studio Discovery workspace, showing a model's embedding space as a coloured cluster map

Squint Vision Studio is a cognitive engine for analysing trained models. It provides workflows for inspecting labelled datasets, quantifying and visualising training results, and examining model behaviour. It can also be used to visualise a model's embedding space. By analysing this space and identifying areas where errors are likely to occur, users can create triggers that improve the model's runtime behaviour.

Instead of treating the classifier as a black box Squint Vision Studio is built to better understand a model's internal feature representation. This space holds a highly informative representation of how a model 'views' samples and makes decisions. By interpreting and understanding how data maps to this feature space, we can improve runtime decision making.

How it works

Import your model

Squint Vision Studio works with trained PyTorch and TensorFlow models. You point it at the model and the dataset it was trained on.

Three stacked planes of sample points, the middle one highlighted

Review the map

Squint Vision Studio constructs a picture of how your model sees its data, where classes separate cleanly, where they overlap, and where mistakes gather.

Converging paths meeting at a node, with a dashed branch leading to a separate circle

Set your thresholds

You decide how cautious the system should be. A high-consequence application sets a stricter bar than a high-throughput one.

A radial dial of spokes around concentric rings and a centre dot

Export the watchdog

Squint Vision Studio constructs a lightweight component that ships alongside your model.

A square linked by a dashed line to a small circle

Route in production

The watchdog labels each prediction as it happens and sends it down the path you configured.

Two overlapping circles with the intersection hatched

From that map, Squint Vision Studio constructs a component that runs alongside your model and checks every decision before it is made. Each one gets one of three labels.