I’m having a blast working with the internal tools Lisa has been building to accelerate our AI work at incident.

In a slightly unconventional working pattern, we’ve ‘ring-fenced’ Lisa to focus exclusively on building internal tools that accelerate the team as they work on products like our incident Chat flow or Investigations.

What this looks like is:

  1. I run a backtest of investigations to target a new feature (finding causing code changes, for example)
  2. Precision/recall isn’t where I’d like it to be, so I try digging into why
  3. Most of the time our tools help quickly finding it, but…

If at this point I can’t easily find the problem, I stop and think “what would be the ideal productised flow for me to debug this?”

I then yeet this feedback into an ai-tooling Slack channel where Lisa normally responds within 60s saying “yep, on it” and several hours later we have the tools available.

So far this has resulted in:

This interaction model may not work for everyone, but it works in our case because:

Honestly though it’s just really fun. The interaction model is high-contact but the pace is great, and the tools are helping us wrestle even our most challenging AI into submission.

You can see a bunch of this tooling in the attached screenshot, if you want to see it in action!