Designing how scientists work with AI in zebrafish experiments

Scientists were already organising every experiment around a physical 96-well plate, a tray of 96 small samples. I used that same structure as the main interaction model in the software.

Before EmbryoNet, researchers often recreated the plate in Excel, printed it, and marked wells by hand while preparing the experiment. Someone rebuilt the digital record afterwards, so a single transcription mistake could carry through to the results.

I designed the setup around the same spatial patterns scientists already used at the bench: rows, blocks, controls, repeated samples and dose series. Spreadsheet import stayed available, but the system previewed imported data against the plate before accepting it.

The microscope captures the same embryo every few minutes. A final image can show what it looks like at the end, but not when a change started, what happened before it, or how quickly it progressed.

I reorganised the analysis around a timeline. Scientists could move through the image sequence while keeping measurements and the model’s classification in the same context. The product stopped treating images as separate predictions and started representing a biological process unfolding over time.

The product tracks developmental time in hours post-fertilisation (hpf).

Once the product tracked time, we could answer a second question: did a higher dose make the same effect appear earlier?

I designed a dose × time view where colour shows the share of affected embryos at each time point. A boundary marks the first time at which at least half of the embryos at that dose were affected. Scientists could then move from the aggregate pattern back to the individual wells and image frames behind it.

Every result still had a confidence score, but asking a scientist to scan all 96 wells meant automation had not removed the review work.

I added a lab-level review rule. The system flagged a result when confidence was low, when it disagreed with comparable wells, or when the system detected something unusual that needed a person to decide. In a typical run, only around 3–7% of results needed detailed review.

Scientists could still open every other result, without confirming each one. Nobody could sign off the experiment until someone had resolved every flagged case.

Before this change, scientists scanned every result. The new review view only surfaces cases that need a decision.

A wrong final result usually started earlier: the model misread the embryo at some frame, and every later read built on that mistake. Overwriting the final label would fix the symptom but leave the time-series inconsistent.

I made correction a two-step flow. First, the scientist finds the earliest frame where the model is wrong. Then they change the result and preview exactly which part of that well will be reanalysed. The scope is visible before anything is committed.

Corrections were useful in two different ways: a laboratory could use approved corrections to improve its own model, and it could choose to share eligible data with EmbryoNet to improve provider models. We made them two separate permissions.

Approved corrections went into a training set the lab owned. A designated reviewer checked them before the lab trained a new model version and validated it against previously reviewed experiments it had not trained on. Sharing approved data with EmbryoNet was a separate opt-in setting.

The system also kept a traceable history of review, corrections, model versions and sign-off decisions.

Scientists used the web application for dense analysis, comparison and model work. At the microscope they stood at the bench, often in gloves, with their attention on the plate and the hardware, not on software navigation.

Reusing the desktop interface would have been the wrong interaction model. I designed the touchscreen around large targets, minimal typing, one task at a time and very clear hardware state.