[throwup238]

I have a similar type of product that depends on a third party and I use LLMs to automate most of the work. I feed it batches of Discord/Discourse messages from the third party’s community along with a prompt containing my feature set and it flags messages that might be relevant so that I can reply to them (manually, although the LLM generates some starting points based on examples replies).

The false positive rate at first was over 50% but with some prompt tweaks and back testing, it’s approaching 10-20%

[kevmo314]

That tool itself might be worth publishing. I've noticed some cofounders (Supabase's kiwicopple comes to mind) that are very on top of HN comments and I am not sure if they're literally reading all the HN comments or doing that in some automated, "notify me when someone talks about Supabase" kind of way. Could be very valuable to notify when even a related discussion happens!

[kiwicopple]

> I am not sure if they're literally reading all the HN comments or doing that in some automated

i use f5bot[0] and have notifications set up for "supabase". it helps that supabase is a unique word so I don't get many false-positives.

[0] https://f5bot.com/

[paradite]

Hey. I actually just built a dashboard and filter for f5bot hits. Would love to know your thoughts and if you think it is useful!

https://tracker.16x.engineer/

[throwup238]

The script is a few hundred lines mostly dealing with the Discord API and Discourse scraping, written entirely by GPT4 so anyone can do it easily. If anyone wants to take the idea and run with it, all the power to them!

HN is easier since you can just use the HN Firebase API. I think people like kiwicopple just use basic string matching, don’t even need LLMs. I only use it because I need to match against my feature set instead of a simple string.

[rovr138]

‘Brand monitoring’ is definitely a thing and there are platforms for it.

You could even start very basic with Google alerts for example.

Still up for new things. Looks like a fun project.