How I Used AI in Division III Player Development

Creating More Time to Work With Players and Strengthening My Player Development Philosophy

---

AI has become the most non-negotiable skill and tool to have across nearly all aspects of life. As it relates to baseball, while much of the AI conversation centers around organizations with extensive resources, I made it my mission to see how we can take advantage and apply it to our process at a small Division III institution. 

In my senior year of college at Stevens Institute of Technology, I was fortunate enough to serve as our Director of Baseball Analytics for our varsity baseball team. As a Division III program, we simply did not have access to the same technology or financial resources. We operated primarily with a Rapsodo unit, a radar gun, and recently installed AWRE cameras that were still being integrated into our player development process. We did not have a TrackMan or any in-game pitch tracking technology (other than a radar gun). We had no dedicated software engineering staff. All we had was myself along with 6 other volunteer students committed to improving our process and reaching our end goal of winning on a national level. 

Rather than viewing those limitations as barriers, we viewed them as opportunities to build efficient systems with the resources available. At least in the beginning, AI became less of an evaluation tool and more of a development accelerator. It allowed us to automate repetitive tasks, rapidly jump start new projects, generate code, as well as troubleshoot technical challenges. Personally, it freed up a lot more time for me to spend with players translating analytical information into a way that makes sense for them, as opposed to strictly time creating and finding the applicable information. 

Player development at the Division III level presents a unique challenge as programs are expected to maximize player performance while operating with significantly fewer technological resources. Every decision regarding technology, data collection, and workflow must be weighed against available time and personnel. Because of these constraints, every project we pursued began with the same question: will this help our players improve in either the short term or long term? If the answer was no, it was not worth our limited time. This philosophy prevented us from collecting unnecessary data never to be used. Instead, each project was designed to either answer a specific coaching question, provide measurable feedback, or improve communication between coaches and players.

Rather than spending hours troubleshooting Excel formulas and debugging Python code, the use of AI significantly reduced development time while allowing analytical decisions to still be made by our coaches and staff. A big message that I kept trying to emphasize within both our volunteer staff, and broader overall program, was that the use of AI was not going to magically make us better. It wasn't going to spit out code to tell us what pitches to throw in what counts, it wasn’t going to fix a pitcher’s mechanics, and above all, I wanted to emphasize that it was not something we were going to blindly trust and solely live and die by. I wanted it to be used as a tool to increase our accuracy in decision making, and shorten the length of time spent on each project. At the end of the day, myself along with the rest of the volunteer and coaching staff, were going to make the most informed decisions possible. The use of AI was just going to help us make better decisions faster and allow us to spend more time communicating with players.

Creating Measurable Feedback for Improved Decision Making

One of our primary initiatives involved developing a live in-season pitching dashboard that tracked both individual and team performance throughout the season.

It started as a simple excel file and each pitcher on our staff had their own individual tab. Every appearance along with their metrics were entered following games. Once new information was recorded, the home page of the workbook automatically updated weekly rankings, season rankings, team leaderboards, and cumulative pitching statistics.

AI played a significant role in formulating the appropriate Excel formulas that powered the workbook. Complex ranking functions, dynamic references, conditional calculations, and automated summaries were developed much more efficiently than they otherwise would have been through traditional trial and error. I simply used Claude and wrote a prompt with my goal, how I want it to look, and what the purpose is and it gave me everything I needed in a fraction of the time it would’ve taken me alone.  

The leaderboard tracked metrics that aligned directly with our pitching philosophy set forth by our pitching coach including

• Strike Percentage

• First Pitch Strike Percentage (FPS%)

• Whiff Percentage

• Maximum Fastball Velocity

• Additional internally tracked development metrics

The workbook also calculated weekly team averages and compared them against season-long performance, allowing coaches and players to quickly identify whether the staff was improving in the areas we emphasized most. More important was that every pitcher had access to these leaderboards. The purpose was not to compare talent. It was to create accountability.

We wanted to reinforce the idea of competition amongst our staff. By showing where they ranked amongst the rest of the team it motivated them to either push harder to climb the leaderboard or remain at the top. By providing transparent, objective measurements of performance, players established measurable goals while creating healthy internal competition throughout the staff. Weekly improvements became visible, trends could be monitored, and both coaches and players gained a clearer understanding of where development was occurring and where it was lacking.

The system also allowed coaching decisions to become more proactive. Declining strike percentage, a falling first-pitch strike rate, or a reduced swing-and-miss ability often became visible, and more importantly, measurable. Rather than reacting after on field performance results had deteriorated, adjustments could be made much earlier on before they drag on for a longer period of time.

Why I Chose Manual Data Collection

One improvement to this to make it even more technologically advanced was to have data pulled from a D3 Baseball Data API. I could have also implemented web scraping to pull data directly from the Stevens website. Either of these certainly would have been much easier than manually inputting data; however, I intentionally chose not to implement either approach.

First off, there is no official NCAA D3 API, so I would need to find a trustworthy source to use an API. As it pertains to web scraping, official statistics at the Division III level frequently vary depending on individual scorekeepers at the school. Subjective decisions regarding hits and errors can differ significantly between games, and statistical attribution occasionally contains inaccuracies between players. They also don’t keep track of every metric that we wanted. 

Maintaining our own data ensured scoring consistency throughout the season while allowing us to record information unavailable through publicly available statistics. Although manual entry required additional work, consistency ultimately proved more valuable than automation. Accurate developmental data was much more important to me than faster but less reliable information.

Translating TrackMan Data into Player Development

One of the more valuable opportunities throughout the season came during a road series at Catholic University. Unlike our home facility, Catholic operated a stadium TrackMan system during games and was kind enough to share the complete pitch-level data following the series.

Since in-game tracking data was not something our pitchers typically received, this provided an opportunity to introduce them to higher level analytical reports.

Using Claude again, I generated Python code capable of producing individualized visualizations from the TrackMan data. Although I know R is the more commonly used language for visualization, I was significantly more comfortable working in Python as it was the language I had primarily been using throughout my classes. Because I was more comfortable with it, I was able to make my own adjustments more so than if I had used R which is why I chose Python.

Each report included visualizations covering:

• Pitch usage

• Velocity distributions

• Movement plots

• Release plots

• Pitch location chart

• Chase percentage

• Whiff percentage

However, creating the visualizations represented only part of the process. Because the TrackMan operator was understandably unfamiliar with our pitching staff, certain pitches were misclassified. Before sharing reports with players, I manually reviewed every outing and corrected pitch classifications to clean the data as much as possible.

AI accelerated report generation while my baseball knowledge ensured the reports were accurate. Perhaps more importantly, it allowed me to spend more time explaining to our players what each area means, and how we can use it to improve, reinforcing the idea that AI was just used to speed up our processes while still leaving ultimate decision making to us. 

Many of our younger pitchers had never previously seen movement profiles, release point consistency plots, or advanced swing-and-miss data associated with their own pitches. Charts without explanation provide limited value. Helping players understand why their fastball is or isn’t generating swing and miss, how their release consistency influenced command, or why one breaking ball consistently outperformed another became significantly more valuable than producing the graphics themselves. The objective was never simply to provide more information, rather more accurate information that I can use to translate to players so they know how it applies to them. 

Estimating Arm Angle

One project that emerged from our growing collection of Rapsodo and AWRE data involved estimating arm angle at release. Although Rapsodo provides release height and side release measurements, it does not directly calculate arm slot.

Using AI as a research and development tool, we developed an approximation model based on available measurements including 

• Release Height from Rapsodo

• Release Side from Rapsodo

• Estimated shoulder pivot height

The primary approximation involved estimating shoulder height at approximately 75 percent of an average six-foot pitcher's height, resulting in a pivot point near 4.5 feet. While this introduces a small amount of estimation error, obtaining precise shoulder measurements at ball release for every pitcher would not be practical within our environment. To validate the model, estimated arm angles were compared against synced AWRE camera footage. Although not mathematically perfect, the estimated values aligned closely with observed video, providing sufficient confidence for our purpose of developmental analysis.

The purpose extended beyond simply calculating another number. By estimating arm angle, we could compare expected pitch movement based on release characteristics against the actual movement observed through Rapsodo. When movement consistently exceeded expectations for a given arm slot, it often suggested unique pitch characteristics worth preserving. Identifying those outlier traits allowed player development conversations to focus less on forcing mechanical conformity and more on maximizing each pitcher's individual strengths.

Future Projects

Another ongoing project involved researching how a Stuff+ model could be adapted to the Division III level. Public discussions surrounding Stuff+ typically assume access to large datasets, extensive pitch tracking technology, and advanced computational resources. Our environment presented a considerably different challenge. Rather than attempting to recreate professional models exactly, we used AI to better understand existing methodologies, exploring possible implementations, and identifying which variables could realistically be collected within our own program. The objective was never to duplicate Major League models, rather to scale underlying concepts into a framework capable of improving player development with the information available.

Philosophy on Player Development

Perhaps the most valuable lesson I learned from incorporating AI into our workflow was recognizing what AI cannot replace. It cannot build trust with players. It cannot understand individual personalities. It cannot determine whether information should be communicated immediately or introduced gradually throughout a season. Most importantly, it cannot translate analytical information into meaningful coaching conversations. That is where I came in to serve as the bridge between analytical information and the players and coaches.

Every player processes information differently. Some prefer analogies. Others respond more effectively to video, mechanical cues, or simplified objectives. Successful player development requires adapting communication to the individual rather than expecting players to adapt to the analyst. As a coach or an analyst or anyone who may be involved in the player development process, it is our responsibility to extract every bit of talent we possibly can out of the player. It is not their job to find a way to listen to us, rather it is our job to find a way to communicate with them. The same philosophy might need to be explained to 10 different players in 10 different ways and that is okay! It is our job to relay the message in whatever way clicks for the player. Each player has their own way and quirks and it is our job to put them in a position to succeed. 

While I firmly believe using AI improved the efficiency of our decision making process and accuracy of our data, I believe it was more so our ability to communicate what we were seeing that allowed us to succeed. Our communication of such projects is what gave us our competitive advantage.  The time saved through AI-assisted coding and research was ultimately reinvested into player meetings, bullpen discussions, individualized reports, and collaborative development planning which ultimately led to two of our players transferring to high level Division I programs. It was important for me to not use our limited resources as a barrier and a hurdle but as a space for creative innovation.