Two decks, exported for Iran
On 4 May 2020 the AI transformation material went to an Iranian audience as two PDFs, exported three days earlier and carrying the cover slide of a January event in Assen.
The folder for 4 May 2020 holds three files and no photographs. Two of them are PDFs, exported from Keynote at 15:19 and 15:31 CEST on Friday 1 May, three days before the session. There is no venue, no attendee list, no recording. The material went out over a link, and the export was the preparation.
Becoming an AI Company runs to forty-three slides. Five steps, from initiate and align through optimise, automate and transform to improve. Six layers underneath: service, technology, process, people, channels, target operating model. A grid adapted from S. Floricel plots technological uncertainty against market uncertainty. Five organisational models, taken from Gassmann and von Zedtwitz on international R&D, ask where a company should place its machine learning people. Tesla carries the worked example: Project WARP, built in four months by twenty-five engineers rather than bought from SAP or Oracle, and per-camera networks doing semantic segmentation, object detection and monocular depth estimation. The last third is a self-assessment, nine dimensions with a score column and a blank box asking where you are today.
The deck then turns to deepkapha's own work. Project curae.ai reads apical lesions in dental radiographs, with Faster R-CNN on ResNet backbones, U-Net for semantic segmentation of pathological regions and Mask R-CNN for lesion instances, trained on commodity GPUs alongside an Nvidia DGX-1. Project caeli fuses frequent 300-metre satellite images into higher resolution ones, citing Amnesty International's search for destroyed villages in Darfur and fire prediction in Australia as the reason to bother. The cover slide was never changed. It still reads LinkedinLive Event, Jan 31st 2020, Location: Assen, with thanks to Ultraware BV.
The second file is filed as Greece-Keynotepptx, twenty-six slides of older company material. It sets three pillars, AI maturity, AI skills and data-driven ROI projects, then gives a client story for each. The skills story runs eight months at a global telco: a blended training programme, skills reported against an evolution matrix built on a few hundred data points, business plans for more than twenty AI solutions valued at 2.5 billion euro, and a first AI Value Map by business domain. The projects story is a four-month pilot, a breast cytology classifier at roughly 82 per cent accuracy on what the slide itself calls a very small dataset. A management slide lists Tarry Singh, Foke S. van der Helm and Jie Mei.
The two decks disagree about their own vintage. One says deepkapha has trained over 25,000 learners. The other says 20,000, with an ambition of 100,000 by the end of 2019. Who convened the 4 May session is recorded nowhere. Neither deck names a host, a university or a company, and no Iranian organisation appears in either file. The country and the date come from the folder. What survives is the argument itself, exported twice on a Friday afternoon and left in the folder.
