# Hornet nest detection: from over an hour to under 5 minutes

> Takeover of a PyTorch model delivered as zip archives, inference audit, and a turnkey GPU laptop that analyses one hectare of drone photos in under 15 minutes.

- Client: Bees For Life
- Period: May to August 2026
- Stack: Python, PyTorch, Streamlit, Ubuntu, Nvidia GPU, Git
- Outcome: 10 drone photos processed in under 5 minutes, down from over an hour

## Starting point

Bees For Life locates Asian hornet nests on drone photos. The detection model, written in Python and PyTorch, had been delivered as undocumented zip archives, with no version control and no tests.

On the client's MacBook, 10 photos took over an hour. One hectare is about 100 photos, so roughly ten hours of computation. The software was considered unusable in the field, with a demonstration scheduled at a European consumer fair in October.

## The audit

One week of performance auditing was enough to find the main cause: inference was running on a single CPU core. Reconfiguring PyTorch to use every core brought the 10 photos down to about 20 minutes, on the same machine.

## The delivery

The client received a laptop with an Nvidia GPU, pre-configured with Ubuntu and the full software stack: Git deploy token, one-step launch, Streamlit interface.

- 10 photos in under 5 minutes.
- One hectare in under 15 minutes.
- The legacy codebase was rewritten and put under version control, with AI-assisted development (Claude Code).

## Today

The client runs the analysis alone, in the field, and hands land owners a GPS nest report during the same visit.