Non‑coder built an ML pitching platform
A former baseball player built an AI‑powered pitching analytics platform that analyzes 8.9 million pitches using custom machine‑learning models, showing high‑impact, accessible project designs outside formal CS backgrounds. The example was shared as a concrete, replicable project idea for people who want to show domain knowledge plus product‑oriented ML. It demonstrates that large‑scale sports datasets can be mined for portfolio projects that combine domain intuition with ML engineering. (x.com)
A former Major League Baseball pitcher named Robert Stock spent a few months teaching himself enough coding to build a public pitching model that had usually lived behind team walls. He said the tool grades pitch quality, location, and arsenal fit instead of just showing a radar-gun number. (outkick.com) Modern baseball already runs on machine data before it runs on box scores. Major League Baseball says a single game now generates more than 15 million data points, and its Statcast system has tracked pitch and ball data across all 30 parks since 2015. (cloud.google.com, cloud.google.com) That matters because a pitch is more than speed. A 97 mile-per-hour fastball and a 97 mile-per-hour fastball can play very differently if one has more spin, more ride, or better placement at the top of the strike zone. (baseballsavant.mlb.com) Teams have built private systems around that idea for years. Stock’s pitch on X was that one person without formal software training can now build a version of that stack with public data and modern machine-learning tools. (outkick.com, x.com) Stock’s own project is called Stockyard Baseball, and his public support page says it uses 7.7 million Statcast pitches across 11 seasons. That page says the grading models were trained on about 5.7 million pitches and now power profiles for more than 2,500 pitchers. (ko-fi.com) The point of a model like that is not to replace a coach with a robot. It is to turn a giant table of release points, spin rates, and movement numbers into a simpler answer like which pitch plays off your fastball and which one gets hit too hard. (ko-fi.com, outkick.com) Baseball already has a market for tools that do this translation. PitchGrader says its older system has analyzed more than 4.2 million pitch shapes and was used by 30 percent of Major League Baseball teams, while its new web product promises advice for individual players who do not want to become data analysts. (pitchgrader.com) What makes this story travel beyond baseball is who built it. Stock was drafted as a catcher in 2009, converted to pitcher later, reached the major leagues in 2018, and is now using playing experience to decide which questions the model should answer first. (wikipedia.org, baseballsavant.mlb.com) That is why this works as a portfolio blueprint for non-computer-science people. The hard part is often not inventing a brand-new algorithm, but choosing a real dataset, shaping it around a real user, and packaging the result so a coach or player can act on it in one sitting. (x.com, ko-fi.com) Baseball is unusually good for this because the public data is huge, the users are obvious, and the feedback loop is fast. If you can explain why one slider tunnels with a fastball and another hangs in the middle, you are showing product sense, domain knowledge, and machine-learning judgment at the same time. (baseballsavant.mlb.com, mdpi.com)
