I hardened up a Postgres database enough to give the public access to run arbitrary queries over an interesting and growing dataset.

I've provided a Claude-optimized prompt for driving this search experience. Opus 4.5 is great at writing SQL and taking advantage of our other features.

There's over 23M embeddings (2048*32 bit Voyage-3.5-Lite over ~300 token chunks, 20% overlapped), 600 GB of indexes for performance, a prompt for letting Claude Code w/ Opus 4.5 drive usage, and there's full support for vector algebra. One can actually embed arbitrary concepts like "FTX" and "guilt" vectors and store these embeddings on the server, and refer to them in queries like @FTX + @guilt.

Because you and your agent can write and run 100k character SELECT statements for up to 60s, you can really answer some incredibly nuanced questions very easily now.

I'm doing a startup building subjective judgement infrastructure, HMU if interested in angel investment.