Updates January 12
PGrid
- Got RAM classification working.
- Added extra steps in variant specs search: part numbers are expanded via a Gemini search alongside the usual single-attribute validation.
- Updated the specs search runner to share more code across categories.
- Added an ad hoc deduplication step for part numbers: it groups variants that share part numbers and asks the AI to search and select the correct variant.
- Added a cron job that checks for new deals across all categories using the ClickHouse “new deals” query used by the internal CLI.
- Added a pipeline health TUI for checking and clearing classification pipelines and specs queues:

- Added RAM to the Next.js frontend, including ClickHouse queries, table, and details pages.
- Refactored to share a concurrency utility across specs search and classification code. There are lots of places where we can run LLM calls concurrently.
- Wrote an ad hoc runner to verify specific attributes. Used it to verify (or add) TBW for all SSD models, then display that info in the frontend specs.