AREA (Automated Research Evaluation Assistant) answers questions about industry payments to doctors from the public CMS Open Payments record, with every number cited to a real row. A customized AREA system was delivered for a fintech platform. The public version below shows the research approach on a separate public dataset.
/api/area-ask. Two calls to AWS Bedrock (Amazon Nova Lite, us-east-1): one writes the SQL, one writes the answer.Loading years covered…
Ask about industry payments to doctors, by manufacturer, specialty, state, or quarter.
Eight questions have a reference query written by hand when the build shipped. Ask one and the receipt shows whether the agent's answer agrees with it.
Every question runs a live, read-only query over the real, loaded rows. An answer with an uncited number is flagged.
Latest scheduled check: loading…
Open: three share questions differ from their reference query, and the first-quarter 2021 question returns an error. History on Live Checks.
The retrieval, the guardrail, and the citation checker: github.com/alphan-ml/area
Each question becomes a real SQL query against the payments table that's actually loaded. Nothing is answered from memory or estimated.
A citation checker walks the answer and matches every number to the row it names. A number without a backing row is flagged as unverified. A cited answer can still differ from the reference query; the receipt shows both checks separately.
The record is loaded year by year. The page always says which years are in and which aren't, read from the database, not typed by hand.