Ask HN: What LLM apps have you built inside your companies? |
Ask HN: What LLM apps have you built inside your companies? |
We have several modules which have a lot of logic, but which requires the user to know what to do.
Often users have been poorly trained when there's been staff changes, and they're not using those modules often.
We do have help pages and a manual, but it seems some people struggle to utilize those resources. Using help pages and manuals to solve your very specific problem is a skill, and not one that's taught at most schools, so this is understandable.
Thus I've been considering if LLMs could help by providing a more free-form input where the user could describe their needs or issue and the LLM could suggest which operations to do and in which order, which data must be entered where and such.
One obstacle is that ChatGPT and friends seem to be very tuned to always provide an answer. In the help case, I'd want it to ask clarifying questions if there are certain variables which must be determined.
For example, consider a LLM front to the typical disk partitioning and formatting dialogs. Determining which file system to use is important, as it may matter to the user. So if the user doesn't specify it in their original query I'd like the system to ask rather than just provide steps for a specific one.
Made because in my previous squad I was basically told my UIs are terrible lol.
[0] - https://chromewebstore.google.com/detail/ui-copilot/hgaldpfd...
I'm just starting using CoPilot and persistence to build RateMyPhotos[1], a python3 script that loads and runs aesthetic-predictor locally against your photos.
Long ago I had a lot of photos up on Flickr, then I tried to prune it using the likes from everyone else, and lost all of the likes... and then ragequit.
That was one of the stupidest things I've ever done. Since then, I've signed back up[2], but there's no way I'm going to manually go through 375,000 photos again and pick out the best ones to upload again.
I expect it to take somewhere around 3 seconds per photo, in about 2 weeks, it should be able to score all of them.
1. A RAG over all the data in database and other stores 2. Summarising information for our customers on what data they can see.