Hello. I'm Daniel.

In 2013 I detected right-whale calls in hydrophone recordings by turning audio into spectrograms and using a convolutional network designed for images. The model did well in Kaggle's whale-detection challenge. Then researchers at Oregon State University supplied much messier continuous recordings.

My shiny model was lousy on them. I had some listening to do. The competition examples had been preselected by another detector, so many confusing sounds were missing from training. I started again with a few hundred examples. I listened to the worst mistakes, labelled them, and fed them back into training. Then I repeated the process. The whale post explains what happened.

PhotoID began because I'm bad at identifying birds and wanted to learn their names. I built the recognition backend and public Twitter bots for birds, butterflies, and wildflowers. Asked to identify a moorhen, it answered mallard. Its second guess was moorhen.

In 2009 and 2010 I volunteered with Python African Tour, working with local organisers and fellow trainers. I co-taught Python and Django in Dakar, Senegal, and led the absolute-beginner group at the African University of Science and Technology in Abuja, Nigeria.

My beginner material wasn't quite beginning at the beginning. I adapted it with Turtle graphics and a Hangman game so learners had something concrete to make and change. The Python 2 tutorial shows how the course introduced functions, loops, and objects. Its examples have not been updated for Python 3.

At 4Catalyzer I worked with doctors and product teams on guidance intended to help non-experts acquire diagnostic-quality ultrasound images. I customized a U-Net and worked on data acquisition and annotation systems.