Baseball / structured data
Groundball
A source-grounded baseball assistant hosted on Cloudflare. Ask supported baseball questions, follow up on an answer, and inspect the queries and sources behind the facts.
From question to evidence
- Ask a supported question. The assistant interprets the request and selects supported queries or source lookups. Structured query recipes are also available.
- Validate the plan. A closed query-plan format limits which operations can execute. Arbitrary SQL and formulas are rejected.
- Retrieve the facts. Historical statistics come from parameterized DuckDB queries. Supported current information comes from bounded source lookups.
- Inspect the answer. Facts carry citations and source evidence; statistical query results retain the SQL, bound values, and source fingerprints.
The model helps interpret questions; the application validates the requested operations and constructs factual answers from checked evidence. The underlying deterministic query engine also works locally without an LLM.
A question to try
Who had the most RBIs in 1962?
Try it in the public app, then inspect the SQL and source evidence behind the result. The repository also includes local setup instructions, a query CLI, and a web interface.
Project scope
This is self-directed, AI-assisted engineering work. Query support is defined by the project's catalog; it is not a promise to answer every baseball question. The source includes tests and release checks for query behavior and evidence integrity.
The public app needs no account or connection to my Mac. It supports historical statistics, reviewed stat definitions, and selected current information such as batting leaders and probable pitchers. Missing evidence and unsupported requests are made explicit.
Supported requests can combine a few operations and use follow-up context. This is a bounded assistant, not an autonomous research agent or a promise to answer every baseball question.