Updates September 29

PGrid

updated scraping screenshots feature flag to handle slightly more complicated screenshot config. Only works for sharded Amazon scraping at the moment.
reviewing SSD model classification with the help of both Claude and GPT-5. Working on a set of titles to use for my own eval script which I can use to evaluate different models.
- started working on an eval utility to compare models easily on difficult examples but this is not complete.
updated Amazon USA scraper to handle cases where Amazon doesn’t register Amazon proxy correctly as USA (saw in screenshots it was Canada sometimes) and updated scraper to handle these cases by clicking through menus to set its location to USA. Not deployed yet.
started working on fullscreen tables which will replace the main page tables to allow main page to be a summary of all categories and a deals page. Incomplete.
more prompt updates for model classification. It is still producing a lot of incorrect classifications but it seems that new Claude Code (Sonnet 4.5) and GPT-5 can pick up these incorrect classifications using the CLI to export a classification session into a text file.
- fixed bugs with model classification where if there is an error with the review step of model classification, it just uses the results of the initial classification.
- also fixed a bug where it would set a classification ignore reason to an item.
created an Instagram account for pgrid_au and posted a first meme reel.
migrated SSD model selection BAML prompt to have a decision output like variant classification does. This required a bit of updates to test files.
- from the migration found that the current model stage was quite poor code quality due to a large amount of “effects” being generated. These were effectively little objects that were later parsed to do some database operation, but there were too many and it was done in a poor way. This led to a new rewrite of the model stage and how effects are done. This took the rest of the week without being completed.