Skyline

Rough high-level overview

Inspiration

The inspiration behind Skyline was to bring the excitement back to baseball by helping fans not only watch the game but truly understand it. While the atmosphere of the ballpark, snacks, and camaraderie enhance the experience, nothing beats the joy of grasping the strategies behind each play. By addressing common frustrations expressed on platforms like Reddit, where fans often struggle to follow complex decisions, Skyline turns the game into an engaging story, encouraging fans to discuss and celebrate pivotal moments with confidence and clarity.

What it does

Skyline is your ultimate companion for understanding baseball. It provides:

Real-Time Strategy Insights: Simplified explanations of game-changing plays and strategies as they unfold.
Key Analytics: In-depth analysis of pre-game lineups, batter-pitcher matchups, win probabilities, and injury reports.
** Casual Mode**: Intuitive breakdowns of gameplay for casual fans to build their understanding.
Predictive Features: Advanced forecasting of team win probabilities, player performance (e.g., home run likelihood), and high-impact plays.

Skyline bridges the gap between casual viewers and die-hard fans, fostering deeper connections to the game and transforming every viewer into a confident commentator.

How we built it

Building Skyline was a multi-faceted effort that combined machine learning, cloud computing, and mobile development to create a seamless and insightful user experience.

To create the Win Probability Impact and Key Moment Analysis feature, we started with Google Cloud services and historical Retrosheet data. Using BigQuery, we processed play-by-play datasets to extract critical features like inning-wise scores, run differentials, and defensive metrics (e.g., strikeouts, home runs, double plays). Derived features such as win probability and game phases (early, mid, late innings) were calculated using SQL queries. Once the data was cleaned and structured, we exported it into Vertex AI, where we trained a regression model to predict win probabilities based on play events. High-impact plays—moments causing significant win probability shifts—were identified and highlighted for users.

For the batter-pitcher matchups, we again utilized BigQuery to process historical Retrosheet data. We calculated key statistics such as batting averages (e.g., singles, doubles, home runs) and situational performance metrics like success in “high-pressure” zones. These statistics were stored in BigQuery tables and used to train a regression model predicting a batter’s success rate against a specific pitcher.

To power simulated Real-Time Strategy Insights feature, we fine-tuned the Gemini Pro model on Google Cloud’s Vertex AI. Using Retrosheet’s play-by-play event code guide, we prepared training data where user roles prompted explanations of play codes, paired with responses that broke down the plays in natural language. The fine-tuned model was then deployed in Vertex AI, enabling it to dynamically interpret and explain play strategies during game replays. Along with data from Retrosheet we used the StatsAPI to get PLayByPlay data and extract Batter, Pitcher, and Field Stats from the API to provide context around the play.

The backend game replay logic was hosted on Cloud Run, leveraging Python and Flask to stream curated insights to users. Firestore was used for managing state and user preferences, providing a dynamic and personalized experience. Our user interface was implemented in React Native, allowing us to build a cross-platform application for both iOS and Android users. We integrated Firebase for user authentication, ensuring seamless access and personalization.

We used Google Imagen to generate our app logo

By combining cutting-edge Google Cloud tools with thoughtful mobile development, we created an application that delivers meaningful insights to fans of all levels, transforming the way they experience baseball.

Challenges we ran into

Diving into this project presented some steep learning curves for both of us. Neither of us had much experience with React Native or Firebase, so we had to hit the ground running—learning to build a mobile app and implement Google authentication as we went. It was a crash course in mobile development and cloud services, but we embraced the challenge head-on.

Budget constraints were another hurdle. Working within a tight timeline and limited resources meant we had to be strategic with every decision, especially when it came to training and deploying machine learning models. We found ourselves constantly juggling priorities, balancing costs, and making the most of Google Cloud's tools to stay on track.

Lastly, the skill gap in machine learning was something we had to overcome together. Neither of us are data scientists or ML engineers, but our curiosity and determination to learn carried us through when we felt like giving up. Tackling tasks like training, evaluating, and deploying models became an opportunity to grow beyond our comfort zones.

Accomplishments that we're proud of

Despite the challenges, we’re incredibly proud of what we achieved. Submitting our MVP within the hackathon deadline felt like a huge win, especially since we managed to balance this project with full-time jobs and other personal commitments.

What we learned

This project was a massive learning experience for both of us. It was our first time using Firebase, and figuring out how to manage user authentication and app state was a steep but invaluable learning curve. We now feel confident in leveraging Firebase for future projects.

On the machine learning front, training, evaluating, and deploying models using AutoML on GCP was an eye-opening journey. Integrating these models into our application gave us a hands-on understanding of how machine learning can enhance user experiences.

Additionally, diving into mobile development with React Native taught us how to build cross-platform apps from scratch. This experience was not only rewarding but also a skill set we’re excited to carry forward in future projects

What's next for Skyline

One key area we’re excited to explore is enhancing our predictive analytics, and introducing even more detailed player and team performance metrics. We also want to make the app more interactive, with features like live fan polls and strategy quizzes to deepen engagement.