Google Cloud's home run predictor at the 2025 MLB All-Star Game

What the system actually does at Truist Park

Google Cloud built a tool with MLB's Statcast unit that estimates where a potential home run might land in the stands during the All-Star Game. As players come up to bat, the model predicts the most likely landing section if that batter hits a home run.

The prediction rests on historical data across the full All-Star roster: conventional stats like batting average and home run percentages, plus data on the direction of hits from stadiums across history. Weather is a factor too — the source names wind direction and temperature as inputs.

In a pre-game demo, the model determined that a Shohei Ohtani home run would most likely land around sections 152-154. That specific output shows the system produces a concrete seat range, not just a probability score.

Two models doing two different jobs

The interesting design choice here is the split. A predictive model handles the trajectory estimate and picks a section. Only after that does the pipeline hand off to Gemini 2.5 Pro, which generates a few dozen candidate messages for the jumbotron, scoreboards and concourse screens.

This is a sensible division of labor. The generative model is not being asked to predict physics; it is being asked to write copy once a section is chosen. Google credits Vertex AI with orchestration, and notes the billboard version returned messaging in a matter of seconds.

Human in the loop is not a footnote

The source is explicit that during the game a team of human reviewers would select a favorite message and tweak language for length, word choice or layout before it appeared across the park. Google names this practice directly.

If this was during the game, a team of human reviewers would select their favorite option, quickly making any tweaks to the language if necessary, whether for length, word choice or layout.Montana Labs

That review step matters because the generated text is going onto stadium screens in front of a live crowd. The model produces options; a person makes the final call. It is a reminder that generating a few dozen candidates is cheap, and the value of the system lies partly in giving reviewers a fast slate to choose from.

The billboard variant reveals the harder engineering

Before the game, the same engine powered mobile billboards driving around Atlanta with location-specific messages — for example, a line about a ball off Ronald Acuña Jr.'s bat traveling 121 MPH, framed against average traffic speed in the city.

Google describes the billboard version as more complex than the in-stadium one: it expanded beyond home runs to pitching and other stats, and had to fold in location and traffic data alongside weather. The added inputs, not the novelty of the idea, are where the orchestration work concentrated.

What a stadium demo tells you about agentic pipelines

Stripped of the event glamour, this is a two-stage pipeline: a domain model that narrows a prediction, a generative model that produces bounded text options, and a human gate before anything ships. Google frames it as agentic AI alongside examples like Gemini calling a restaurant or banking customers building research agents.

The practical lesson is that the demanding parts were data plumbing and latency — combining historical stats with live batting orders, weather, and traffic, then returning messages in seconds. The specific implication for teams building similar systems: the model choice is the easy decision, while defining a tight generation task downstream of a reliable predictor, and keeping a human review step in the path, is what makes the output usable in a live setting.