Data-Butler: Language comprehension without losing control

How many competition shots has an athlete taken this year? With Data-Butler, coaches can easily ask questions like this in everyday language. A language model understands the question, and verified program code calculates the corresponding number. This makes sports data accessible without technical expertise, and every answer remains reliable and transparent.

Competitive sports organizations constantly generate structured data: training loads, competition results down to individual shots, and daily metrics such as sleep or perceived exertion. This data would be valuable for training management, but accessing it requires navigating interfaces, rigid dashboards, and data exports. Anyone who wants to answer a simple question – such as how many official competition shots an athlete has taken this year – needs technical expertise or the help of others. As a result, data collected at great expense remains unused in day-to-day training.

The obvious solution would be to ask an AI chatbot the question and have it calculate the answer directly from the data. Language models understand everyday language surprisingly well. However, the numbers they calculate themselves cannot be verified: they sound plausible but may be incorrect, and there’s no way to tell from the answer. This is not sufficient as a basis for decisions in competitive sports. The «Data-Butler» project is therefore designed to combine both: understanding freely formulated questions and having the corresponding figures calculated by verified program code.

The «Data-Butler» is designed to allow coaches to query sports data in everyday language and rely on the figures provided. To achieve this, the system strictly separates the process of understanding the question from that of calculating the answer, while also protecting the athletes’ data.

Advantages of the Data-Butler:

Easy access to your own data

• Questions instead of interfaces: no technical expertise required
• Retrieve key metrics directly, without dashboards or exports
• Multi-part conversations with follow-ups and clarifications are also possible

Reliable figures

• All values are generated by verified program code, not by the language model
• The calculation process is documented and remains traceable
• The same query and the same data yield the same result

Honest handling of uncertainty

• The system shows how it understood the question
• If a question is unclear, the system asks for clarification instead of providing a guessed number
• Missing information is flagged as a gap, not filled in

Protection of personal data

• A central rule in the code controls how much data context a model call sees
• Four tiered access levels, ranging from the query text alone to individual data rows
• Individual data rows are only displayed after explicit approval by the user

Long-term goals include:

• Implementation on the Force8 platform with approximately 9'500 users, including about 2'500 coaches
• Expansion to additional sports and associations on the platform
• Establishment as a trusted tool for data-driven training management in Swiss elite sports

Participants

The project was carried out by the Institute for Data Analysis, Artificial Intelligence, Visualization, and Simulation (DAViS) as part of a bachelor’s thesis in collaboration with the industry partners Swiss Shooting and Force8.