Follow-up: what I learned scaling a SQLite/FTS5 patent database from 3.5M to 5.36M records

A few months ago I posted about classifying 3.5M US patents with Nemotron 9B on a single RTX 5090. This is a follow-up on the data-engineering side. Disclosure up front: I'm a patent lawyer who started coding in Dec 2025, and I build and run the site mentioned at the end (free, no signup, no ads — it's a hobby project). Things that might save you time: - PatentsView moved to the USPTO Open Data Portal in March 2026. The old S3 download links are dead. Abstracts also got split out of g_patent into their own