Research becomes companies.

It read apical lesions in dental radiographs: Faster R-CNN on ResNet backbones to find them, U-Net to segment the pathological regions, Mask R-CNN for individual lesion instances, trained on commodity GPUs alongside a single DGX-1. The same line of work put the lab in the ITU and WHO focus group on AI for health, where the dental reporting checklist became its most-cited paper.

It taught, too. In September 2019 curae.ai ran a three-day Deep Learning in Healthcare bootcamp with the computer science department at UT Dallas, two days on convolutional and generative networks and then a medical imaging hackathon. In May 2020 NVIDIA put it on screen in the GTC keynote, among the researchers using its hardware against COVID-19.

caeli did multi-image super resolution. Satellites revisit often but see coarsely, so it fused many frequent 300-metre images of the same ground into one sharper picture, working on the European Space Agency's Proba-V challenge. The lab's own slides gave two reasons to bother: Amnesty International looking for destroyed villages in Darfur, and fire prediction in Australia.

It sat with PETAI in the lab's earth sciences and climate cluster, the same ground the EarthScan line grew from, though caeli itself looked up rather than down and did not carry through. It is here because it ran, not because it shipped.