Drug-Safety Agent on Databricks · Status
What is done, what is not, and the order
Of the 35 steps in one investigation, 28 are built, 6 are partly built and 1 are not built. The order below closes the most gaps first.
| # | Next piece of work | What it closes |
|---|---|---|
| 1 | One test per owed harness capability, or an accepted gap with an owner and a date | The harness gate, now 2 of 76 |
| 2 | The open-question case: what should I look at for this drug?, built and tested | The one place a model is genuinely needed; today's agent is mostly a deterministic core |
| 3 | Move the store to a Delta table; put a login in front of the desk; notify a reviewer | The two stand-ins in the desk |
| 4 | Write three hand-off reasons, the freshness window and the erasure rule into the specification | Findings from building and running it |
| 5 | Wire the agent to the live Unity Catalog functions; retire the Python copy | One definition of each tool |
| 6 | Run the 20-pair evaluation on the Databricks model endpoint | Evidence for the 'built on Databricks AI' claim |
| 7 | Stronger scenario checks that also test what the reply said | Scenarios that cannot pass by only calling a tool |
| 8 | MLflow tracing and the shared evaluation library; retire the old first-draft loop | Observability; dead code |
What the demo can honestly claim. A working, tested, rule-bound drafting assistant over public data. A weekly job picks the pairs worth a look; drafts wait in a queue; a named person decides under their own name, and a decision is final; a case the agent cannot finish goes to a person with a reference. Every number in a draft is read from the data. What it cannot yet claim: that it runs on Databricks end to end, that the desk is secured, that the store is the production one, or that the model's open-question ability has been tested.