Updates February 16

PGrid

• Ran a migration for the new GPU classification and generic item table usage, and deprecated the old GPU-specific tables for scraping and classification. After letting the system run for a bit, it seemed like the migration completed OK.
• Did a large clean-up of code related to the old GPU classification and scraping.
• Added a new marketing folder including a CLI. The CLI is for AI agents to be able to do things like use BAML. First goal is to get AI to query xAI to find popular tweets about computer building and see if there can be content ideas from that. After testing the xAI API, I found it was pretty bad at finding popular tweets. Tried to get the OpenClaw agent to go on grok.com but was blocked. Tried a few different ways to bypass but was unable.
• Got an opencode bridge working for BAML to be able to use models configured in opencode. This lets me use Kimi Code from BAML.
• Updated the opencode bridge to allow LLMs to use tools to get batches. This makes a single LLM request able to handle multiple batches in theory. Not tested on evals yet.
• Added a Newegg scraper. Enabled it for all categories currently being scraped from other websites.
• Updated the audits internal CLI command to support GPUs and fixed some minor issues with the TUI.
• There’s a big backlog of items to classify. Currently using Gemini 3 Pro but this is eating up too many free trial credits. I ran evals on MiniMax M2.5 but performance was poor. Gemini Flash is better but misses minor things. Tried to eval the new Qwen model but ran into too many request issues.
• Fixed a bug with SSD classification where it was classifying fewer items per batch than other categories.
• Rewriting email alerts to use ClickHouse and work across all categories.
• Clean-up and restructuring of ClickHouse tests.