An AI analyst that reads Linear, Slack, and GitHub and answers team queries off our dbt models

At Augment Code, like at many other companies, the human data team was the bridge to the data. Every question came through us. The trouble with being the bridge is that you're also the bottleneck, and ours backed up constantly. Context got lost jumping between tools. Definitions morphed without documentation. New people never knew where to look, and most questions were too small to be worth dropping into a queue, so they just never got asked.

Everybody wants self-serve data. We did too. We just kept chasing it with more dashboards, more enablement, and more tooling training, which is a funny way to get people to do things for themselves.

So we built something else.

One coworker with one job

Instead of another dashboard, we built an Expert. An Expert is basically a coworker with one specific job. Ours is a data analyst. We call her Mimi, after Mimir, our dbt repo, which borrows its name from the old keeper of knowledge in Norse myth. A fitting namesake for an analyst, though if we're honest we mostly picked it because it's easy to spell.

The thing that makes Mimi useful is context. She's wired into engineering, marketing, sales, and all of our analytics. She reads Linear, Slack, and GitHub. She pulls from Hex, so her answers come off our certified reports instead of something she made up. And she only has access to what she's supposed to, which matters more than it sounds.

You ask in plain English, she answers. That's the entire thing anyone has to learn.

Ask her anywhere

In Cosmos you open a session and type whatever you'd normally have Slacked the data team about: what is our most profitable product line? You get an answer back. You don't have to know where it's coming from or what a semantic layer even is. We built the skill underneath so the answer always pulls from the right place.

You don't even have to know Mimi exists. Start in a generic session and Cosmos points you to the right Expert, one click and you're there. A new hire doesn't need any tribal knowledge. They just need a question.

01 · Routing — Ask in any session. Cosmos hands you to the expert with the right context, the Mimir SQL Analyst, which runs the query and answers.

01 · Routing — The Mimir SQL Analyst runs the query and answers in the same session.

She's in Slack too. Drop Cosmos into a channel and you can say “Hey Mimi,” watch the little magnifying glass while she works, and get the answer right there in the thread where everyone's already talking.

She works the way we work

Mimi doesn't answer the way a generic AI assistant would. She answers the way our team actually does the work. Our privacy rules and caveats are written into her guides. She follows an order of operations, check this first, then that, then the next thing, the same steps I'd want an analyst running before they hand anyone a number.

She also checks her own work before sending it. And then she still gives you a thumbs up, thumbs down, or an “audit this again” button, which forces a half-second of “wait, is this actually what I asked for?” That little pause catches a lot.

And because the whole exchange lives in a thread, you can keep pulling the string. Ask a follow-up and she carries the context forward, so “how did that change this week?” lands without you re-explaining what “that” was.

Ask in plain English, watch Mimi work, then keep pulling the thread. The follow-up carries the context forward on its own, and one tap sends her back to re-check, where she either confirms the answer or flags the caveat that matters.

We can see everything now

Every question, no matter where it came from (Cosmos, Slack, Linear), lands in a log we can watch live. We see what was asked, what went back, and who asked it.

It sounds like surveillance and it really isn't. It just means that when the AI gets something wrong, we can catch it before it spreads. No more mystery number turning up in a customer deck that nobody on the data team has ever seen. We get the speed and the self-serve without losing track of what's going out the door.

03 · The log — Asked in Cosmos, in Slack, or in Linear, every answer posts to one channel the data team watches in real time.

What changed

1. Having it all in one place is what made it possible. The context, the AI, and the security come built into Cosmos, and the data already lives right there, so we're not piping it out to a separate tool to pull this off. The setup we had before couldn't have supported this — the pieces were scattered across too many tools.
2. Self-serve used to be a full-time job in itself. We always wanted it, that was never the issue. The issue was the path: endless enablement, tooling training, and dashboards to build, then keep all of it alive forever. Maintaining the self-serve machine quietly became the job. It should have been the foundation, clean dbt models, a tight semantic layer, no legacy definitions hanging around in Hex, the warehouse in good shape everywhere Mimi reaches. Now it can be.
3. Speed is what unlocks curiosity. A question you're only mildly curious about is never worth a spot in someone's queue, so you drop it. Once the answer takes seconds instead of days, the small questions start getting asked. Turns out the small ones are most of them.
4. We can finally see what people actually want to know. The patterns, the gaps, the questions that keep coming up. We build for that now instead of guessing at it.
5. The logic lives in one place, right next to the data and the code. No second semantic layer, no business logic smeared across five tools. We change it once and we're done.