Ask HN: Most interesting tech you built for just yourself?

About 8 years back, I was leading an engineering team which was the escalation path for customer support. We were sitting on a large corpus of support tickets but didn't have any insights. I was amazed when word2vec came out and blew my mind. So I built a language model that trained on support ticket data. I modeled system logs attached to support tickets as an NLP model to predict what was going to fail next and for which customer.

Never made it to prod but was a great tool for me to see where I want to budget my team's time.

This is way before all the LLM and Generator models, but it was such a fun project.

>This is way before all the LLM and Generator models, but it was such a fun project.

That means something more sophisticated has to exist today and should be commercially available. Can anyone explain to what extent companies use this stuff in their interaction with customers, and how successful is it? (Somehow I still see AI still as one-off things people do for fun or AI being used to hype up rather mundane software.)

We have a corpus or database of programs that follow logic but with no simulation, so it represents knowledge to solve a problem yet all we have control over is the parameters (inputs). In this case, input is functional logical content (a program), describing the resolution of corpus details. The model solves its integrated, corporate logic, and our output is an interpretation of that individual program.

Now our task is to swap out this entire database for something like it, but not exactly the same. The output becomes the input to this new matrix. The individual program persists, but everything is the next generation. With a little book-keeping, the programs do our will...

Don't think I quite follow. Is the new program (operating on the output of the earlier program) supposed to reason about why you are seeing the result that you are seeing? Or is it doing more post processing to make the earlier output directly consumable by your corporate systems.

It should exist and would be super powerful considering all the recent advancements in language ML. Here was the mental model of my model: the canonical representation (i.e. a representation after taking out run-time populated fields) of a log line represents the smallest meaningful unit of this "log language": _a word_. Taking this analogy further, an event is a collection of logs that occur together (mostly in order)—just like words spoken together form _a sentence_. Finally, collections of events that occur in close proximity (in time) represent _paragraphs_, while paragraphs occurring in a certain order constitute _chapters_. Using this mental model opens the door to apply all the new AI techniques for text extraction, summarization and generation to extract the semantic structure of any "log language" and then learn and classify behaviors observed at run-time. The eventual objective function is not generation though--it's reasoning with the optimal FP-TP tradeoff on a ROC curve.

I haven't seen anyone do it yet. Maybe companies like Splunk and Elastic will take a lead here. I am happy to engage, advise and contribute if there is an open source project around this. Has anyone else seen something remotely close to this?

Very interesting, thanks!

The new program’s purpose could be to do more post processing to make the interpretation of that earlier program directly consumable (inter-generationally), or it could simply start producing more problems to solve.

Gotcha! That makes sense. I would recommend looking at LangChain though, as it does a good job at modeling multi-stage learning / inference environments.

Integration API: https://thetaplane.com/ai/langchain/api

Inspired by: https://github.com/daveebbelaar/langchain-experiments/blob/m...

There's a YC startup that's doing something along these lines. They recently posted something, don't remember if it was on front page or an "YC startup jobs" email.

Wow... that was quick. Would love to see what results you get.

Mendable.ai?

Nah... Mendable seems to be searching in documentation. This idea is to extract an execution model of every component in your system from the logs of that component. We can think of the prediction step as a complete separate machine model for now.

Anything else comes to mind?