Muscat, and a model made public the same day

The invitation came from CCED at a university in Muscat. The session ran on 23 November 2020, over a screen. Nobody photographed it, because there was nothing in a room to photograph. What is left is the deck Tarry Singh presented, a PDF he exported from it two days earlier at 16:20 CET, and two video clips still sitting inside the file. The deck carries no host mark of any kind. Everything known about who sat on the other side of the call comes from how the folder was filed.

Twenty-six slides, four movements. A short account of what artificial intelligence is, with machine learning nested inside it and deep learning nested inside that, each ring carrying its own equation. Then one case from medicine. Then one from corporate change. Then a first look at work DeepKapha had done itself. The framing slide does the heavy lifting: learning to drive, learning to diagnose, learning to make money. Learning is the word Singh keeps returning to.

The medical case is Afghanistan. The healthcare division of GOPA, 50 countries and 3,500 projects behind it, had breast cancer imagery sitting on a client portal, poorly organised, annotated by German pathologists in a form no model could read, and no machine learning skill anywhere in the building. The slide on Dr Sohaila Niazi is not about technique. An Elle profile had found her first. She left a prosperous life, went back to a war-torn region where young women were dying, and paid out of her own pocket to get women to hospital. The AI work was a four-month pilot. It produced a working prototype that separated malignant from benign cytology images at around 98.5 per cent. The deck says prototype. It does not say deployed.

The corporate case is a telco operating in 50 countries on 15 billion euros of revenue in 2020, walked up five plateaus from AI fundamentals to new lines of business. The numbers on those slides are specific. Sixteen weeks for awareness and planning, excluding the physical facility. Eight months to answer whether an AI department could be built from the talent already on the payroll. Around eighteen months to run a blended training programme, report a skills evolution matrix off a few hundred data points, and write business plans for more than twenty AI solutions worth 2.5 billion euros. An AI Value Map ties each one back to revenue, EBITDA, cost and technology.

The last working slide carries a date. After working through DeepLab, U-Net and DANET, DeepKapha had built EarthAdaptNet for the energy work it called PETAI. The slide says the paper and the code go to arXiv and GitHub on 23 November, and then, in brackets, today. The lecture and the release were the same afternoon. The slide types the year as 2021, but the file it lives in was saved in November 2020, so the year is the slip and the day is not.