Drug-Safety Agent on Databricks · Start
Drug-Safety Agent on Databricks, drawn simply
This is a guided tour of one small AI agent, from the question a person asks down to the lines of code that answer it. The agent helps a drug-safety scientist decide whether a medicine and a harmful side effect deserve a closer look. It runs on Databricks, and every page ends in the real code.
built 28 partly built 6 not built 1 of 35 steps in one investigation
How to read this tour. Every page has the same shape: a plain paragraph first, then a picture, then the real code under it. Boxes in the pictures are clickable and open the page that explains that box. The top tabs are the big sections. The left list shows every page. The Back and Next buttons at the bottom walk the whole tour in order.
This tour says what is not built. Where a part does not exist yet, the page says so in a pink badge, and the status page lists what is next. The largest remaining gaps are moving the store to Databricks and securing the desk.
Where would you like to start?
Four words used everywhere
| Word | What it means here |
|---|---|
| Signal | A drug and a side effect that appear together in reports more often than chance would suggest. A reason to look, never proof that the drug caused anything. |
| Tool | A small, named, read-only function the agent may call, such as count the reports. The agent never touches data any other way. |
| Harness | The shared code around the model: the loop, the limits, the checks. It is the same library for every agent we build. |
| Draft | The agent's only output. A person approves, edits or rejects it. |