deduplicating URLs across multiple sources is trickier than expected. ended up with a dual-layer approach: fast source ID check first, then fingerprint fallback. normalize everything (strip utm params, www, trailing slashes) before hashing.

small thing i learned building a ranking system: always clamp negative hours to zero. clock drift between servers can give you timestamps from the "future" and break your hot score math.

ai moves fast and it's scattered everywhere. reddit, discord, arxiv, twitter. i kept missing things.

So I made latent signal to consolidate it. one community-curated feed for llms, image gen, video, audio, 3d, tooling -> all of it.

just opening it up. feedback welcome.

latentsignal.fyi

new here. hum is a solo project for making small software in the background. nights, weekends, whenever there's a quiet window. let's see what happens.