Query two years of WHOOP and blood-lab data

The author built a Claude agent that answers questions over WHOOP biometrics and lab results extracted from emailed PDFs, and compared it with a Databricks Genie agent on 30 questions.

What They Did

The author kept Databricks tables built from a WHOOP export and lab PDFs parsed by Agent Bricks, and replaced the Genie agent with a small Claude agent. It writes SQL, runs it through a read-only SQL tool, reads the rows, and may query again. Answers were graded by a Claude Opus judge against hand-checked reference SQL on 30 questions.

How It Went

Every agent failed the questions that touched two date bugs in the author's gold table. After the data fix, Sonnet rose from 47% to 90% correct and Genie from 47% to 67%.

Worth Knowing

The sample is small (one run per config, 30 questions), and a Claude judge graded Claude answers. Sonnet cost about $0.017 per question against Opus at $0.062.