Snake Detector: a narrow computer-vision demo with a reproducible workflow

Try it

Open the live demo in a new tab. Upload a photo and get a narrow snake vs no-snake prediction from the current public build.

Note: the demo may take a few seconds to wake on first load while the host cold-starts.

The live demo uses a real-photo iNaturalist-trained Keras model. The model file stays out of normal git history, but the release package mirrors the model, deployment config, held-out metrics, threshold sweep, confusion matrix, training curves, sample predictions, and checksum manifest.

The held-out split is 255 snake images and 1,028 no-snake images at threshold 0.76. The snake recall number is why this page keeps the safety boundary explicit: the demo can miss snakes and is not field-safe wildlife software.

What the demo shows

• Deployment and inference path: a working public endpoint with limited inputs and outputs.
• Reproducibility: the training and evaluation loop is scripted; artifacts and the repo back the story on this page.
• It is not field-ready wildlife identification, species-level reliability, or licensing-cleared training data suitable for commercial redistribution.

How to use it

1. Upload one image through the demo UI.
2. Review the prediction and confidence framing shown in the app.
3. Read the known limits above before trusting the output for anything beyond a narrow experiment.

Most of the value is in dataset hygiene, fixed evaluation splits, confusion-driven review, and refusing to let aggregate accuracy hide weak classes. That discipline transfers directly to larger vision projects where bad predictions fail quietly in production.