ATMAN V2

Having built and trained ATMAN v1 completely myself, on the data most readily available (in this case, broadcast tennis video streamed over the internet), I had a working software suite, capable of handling broadcast video from ingestion to analysis and visualization, but no product yet. I had postponed the most crucial questions ("What need do I cater to? What problem do I solve?") in favor of building something I thought was beyond cool. Rookie mistake, I know. To make ATMAN work as a product in the real world, I had planned to approach tennis academies, sports data outlets, and broadcasters, and see if anyone thought ATMAN interesting. Had I gone through with that in a disciplined manner, I'd probably have learned some much-needed lessons on validating before building much earlier; instead, I jumped at the first opportunity to pivot ATMAN into an on-court, two-camera system when SK Tennis Base (who are just the best — love you guys) offered to let me film and prototype and test to my heart's content. It being the advent of generative AI, I rewrote the pipeline to handle matters of synchronization and two-camera capture, improved some models, and built a desktop app and a phone app. The phone app was especially interesting: it was basically a data scientist in the coach's pocket — an agentic system with access to the processed shot-by-shot data, the film, knowledge bases on stroke mechanics, tennis tactics, etc., and the ability to answer a coach's query ("How did topspin rates change after we made a grip adjustment?", "Summarize Mark's serve patterns on the ad side in matchplay situations") from the data. Since I am not all too willing to let internet strangers use up my OpenAI credits, you'll have to make do with the web app, which is still pretty damn cool.