When a major system incident occurs, there is a ton of online and offline communication going on — or at least there should be. The online part creates a digital audit trail of what happened when and how people reacted, and ultimately how the problem was solved. Jeli launched a few years ago to help companies understand and track incidents. Today the company announced the next logical step in that approach — a beta that brings ChatGPT-like functionality to incident management.
Company founder and CEO Nora Jones says that adding generative AI to the product seemed like a natural extension of what they were doing already. “We’re always on the lookout for things to improve the product, and make it easier for humans to understand what they’re supposed to do in the middle of an emergency, or even outside of an emergency and where to access expertise. So we started playing around with generative AI to help during the incident and then after the incident as well,” Jones told TechCrunch.
For starters, it’s going to help humans understand what happened with a summarizing feature. As she points out, sometimes an incident can drag on for days or weeks, and just being able to get caught up without having to wade through all the messaging can save oodles of time for the people involved. “So during the incident, we have a feature called ‘Catch Me Up’ that takes the conversation that’s happened so far, and summarizes it for whoever is needing to be caught up on the incident,” she said.
In addition, the company is providing a narrative for users to get a sense of the breadth of the incident. “We allow you to generate markers that summarize for you where detection happened, who was associated with it, what they said, where diagnosis happened and where repair happen. And the idea is to direct the humans towards understanding anomalous things about how the event unfolded.”
The company is using the OpenAI API for starters to build this functionality, but Jones says they have trained the model specifically on computer incident management. “We’re leveraging generative models from OpenAI. So we were using ChatGPT, and then we were able to improve a lot of the responses through things like chain of thought reasoning. That has allowed us to classify incident transcripts more readily,” she said.
Jones says to limit the hallucination problem where the model makes up part of an answer, which would be especially problematic in an incident management scenario. They show their work, so a human can check the source material to confirm the data they are generating is correct.
“You can see the summary and all the evidence that it pulled from. I never want to call this ‘an automated post mortem’ or ‘an automated incident review’ because it’s not it’s like it’s a kick starter. And we actually encourage humans to edit what it came up with,” she said.
You can’t query the incident report in the current version just yet, but Jones says the purpose of the beta is to refine what they’ve built, and they will decide how to enhance it based on feedback from users.
The company is announcing the new capability today. Interested Jeli customers can sign up for the waitlist.