Two Months at Caelor: Finding Where AI Actually Helps

I signed my internship contract with a company called Vectors. Somewhere in my second month, Vectors and another company merged and became Caelor. My badge changed names before I'd even learned everyone's.

That merger ended up teaching me more than most of the actual project work. Two teams, two ways of writing things down, two ways of running a project, two different instincts about what "done" means, all suddenly sharing the same room. Nobody handed me a single source of truth to accept quietly. I got to see both sides, ask why things were done a certain way, and build my own opinion out of the pieces that made sense. That was the best accident of the whole internship.

The actual job

Caelor builds apps on top of Jira Service Management, and my job was to figure out where AI genuinely adds value across two of them. There was one hard rule from day one: any AI feature had to run entirely inside Atlassian's own platform. No external AI providers. For enterprise customers, that's a trust badge they check before they buy, and losing it isn't worth whatever the external model gets you.

I thought that constraint would limit what we could build. It turned out to be the most useful filter we had, because every idea then had to survive a question most AI features never get asked: is this actually better than just writing the code properly? A lot of ideas didn't survive that question. We went from 27 candidates down to 6.

What I shipped

The feature I built lets someone type what they want in plain English and get their company's data back, no query language required. The part I'm most proud of isn't the natural language parsing, it's that the AI never gets final say. It suggests a query, a human confirms it, and the manual way of doing things never goes away. That human-in-the-loop constraint was as deliberate as the "stay inside Atlassian" one, and for the same reason: trust has to survive contact with an actual user.

The part I didn't expect

I walked in expecting "here's a task, go sit in your corner." That's not what happened. The first draft I wrote got sent back with a simple note: too much, reads like a report, doesn't sound like you. Fair feedback. I rewrote it.

The rest of the internship was the team event where Slack usernames turned into actual people, long dinners that outlasted the food, visiting Tunisian monuments with colleagues instead of as a tourist, and a running joke I didn't understand in week one and was fully in on by week five. I also picked up ping pong from a standing start: lost constantly at first, eventually got just good enough to lose more slowly.

If I'm honest, I left a few things unfinished, and there are colleagues I only really got to know in my last two weeks. I should have started those conversations earlier. Everything else about it, I'd sign up for again.