Pupavision kept photographing fruit-fly development for eleven days while its AI assistant was logged out, its creators report in a preprint first submitted on November 17, 2026. The model helps adapt the instrument; local software takes the pictures.
For small labs, the useful inference is that AI can help customise scientific equipment without being needed for every observation. Whether those observations produce accurate biological measurements is a separate question.
Joel Sadler, Daniel Graham and Melanie Worley at the University of Virginia describe a 26-day run beginning on June 5 that stored about 3,580 images. This is newly available research about a preliminary deployment, not a November hardware launch or an independently replicated result.
What the assistant contributed
The biological task is to record when larvae become pupae, the stage before adult flies emerge. The authors describe researchers returning to an incubator every few hours, across nights and weekends, to count new pupal cases. A camera can preserve observations for later inspection without someone being there for every photograph.
The team built versions using an Arduino UNO Q and a Raspberry Pi, each connected to a day/night camera, with wall power, local storage and a cardboard backdrop. Claude Code runs its tool-executing software on the board, but the AI model runs in the cloud. A small computer beside the flies does not mean all the AI processing happens there.
The authors credit the agent with practical camera work: tracing dark night images to compressed capture, switching to uncompressed frames, adding time for exposure to settle, and changing a setting associated with live-view flicker. They also say it wrote the annotation interface, analysis workflow and dashboard. These are specific reported contributions, not a measured reduction in laboratory labour.
Why logging out did not stop the pictures
The public capture script contains no model call. It tries to obtain a picture from the local camera stream, then falls back to accessing the camera directly. Its default interval is 600 seconds. A separate service configuration tells the operating system to restart the process if it exits.
Taking the next picture is a predetermined job; diagnosing why a picture is almost black is a different kind of task. That is a concrete example of the distinction between fixed workflows and model-directed agents. The assistant can help change the instrument without having to approve every photograph.
Independence is narrower than reliability. The paper also reports a 51-hour recording interruption during development, and power loss, camera failure or a full disk can still stop acquisition. The public toolkit supports the separation of recording from AI assistance; it is not an independently verified reproduction of the pilot, nor proof that every later safeguard operated during that run.
Why ten-minute photos are not ten-minute measurements
Recording an image is only the beginning. Automatic detection was not reliable enough for the intended workflow, so a researcher chooses a late image, draws boxes around visible pupae and assigns them to vials. The software then works backward through earlier images to estimate when each selected pupa appeared.
That human choice defines what the resulting curve represents: the selected, visible pupae, not every animal in the vial. The interface keeps the image evidence available beside the curve so the researcher can inspect individual events.
There is a second, less obvious limit. The timing analysis uses only images from 08:30 to 20:00, excluding infrared night images because they look different. Calculating from those endpoints gives a 12.5-hour overnight span omitted from the analysis. That is an analysis gap, not a camera outage or a measured timing error.
An event spanning that gap is placed by estimating between the available daytime observations. A picture every ten minutes therefore cannot establish ten-minute biological timing accuracy. The authors have not validated the event times against expert scoring and present the curves as sample workflow output, not a biological finding.
The distinction matters beyond this rig: saved observations, a plotted pattern and a defensible scientific conclusion are different achievements. Our guide to reviewing AI interpretations of laboratory data examines that next step, after the observations exist.
What small labs should budget and evaluate
The software is publicly available under the MIT licence. The paper’s US$45–75 parts estimate is the authors’ figure, not a current shopping quote or a complete operating budget. Model access, setup, upkeep and scientific validation are separate considerations.
As checked on November 18, Anthropic lists individual Claude Pro at US$20 a month in the US, including Claude Code access; regional prices and taxes vary. Claude and Claude Code share account usage limits, and uninterrupted service is not guaranteed. That price does not establish that Pro supplies enough usage for a particular lab.
Nor is automated fly photography new. The Hatching-Box study published in 2025 already described camera-based monitoring with detection and tracking software. Pupavision’s authors have not compared their agent with an agent-free baseline or tested unaided novice deployment. The case supports examining AI-assisted adaptation, not assuming it is cheaper, easier or better than existing equipment.
A lab considering the approach should assess three outcomes separately: whether original images keep arriving, whether the assistant’s changes are useful, and whether derived event times agree with expert judgments. The authors recommend protected copies of original images and reviewed, versioned changes to measurement code, while explicitly saying those stronger safeguards are not yet demonstrated in the prototype.
The practical question is not whether a lab camera can run an AI assistant. It is what keeps recording when that assistant is unavailable—and whether the saved pictures support the measurement the experiment needs. Pupavision makes the first separation concrete; the second still needs validation.
