Engineering blog
How this site is built — notes on the data pipeline, search, and the CMS files behind every Medicare Physician Fee Schedule rate we publish.
Rate limiting an API you bill by the code, not the request
A per-request limit on a bulk-capable API isn't a limit at all — one call can do the work of hundreds. Ours counts what the request actually costs, checks two budgets before charging either, and fails open rather than 500ing during a Redis blip.
When a ZIP code can't tell you the truth
Medicare pays by locality, not ZIP code, and some ZIPs genuinely span more than one. We had to decide what our lookup tool says when the honest answer is "it depends" — and what it should have said a year ago, not just today.
One Redis, two jobs, and a cache on a collision course with the queue
Our public-page cache and our job queue share a single small Redis instance. A routine warming job was on a straight path to filling it, and Redis configured noeviction doesn't degrade — it stops.
What 13 years of CMS spreadsheet headers taught us about parsing by name
CMS isn't an API vendor — they ship the fee schedule as spreadsheets, and the column layout drifts almost every year. We stopped trusting position entirely, and the alias table that replaced it turned into a small history of thirteen years of formatting decisions.
Tuning concept search: what the cosine scores actually said
We shipped embedding-based search over 594 medical codes, then measured real queries against the real corpus. Three assumptions turned out to be wrong.
For what shipped and when, see product updates.