Inside AIntOps: a 90-second walkthrough of where your AI budget actually goes

Feature lists and screenshots only get you so far. So instead of describing the loop — connect, see, get alerted, act — we recorded it. Here's what happens in the 90 seconds it takes AIntOps to go from a pasted API key to a quantified savings recommendation.

Watch the walkthrough

What you're watching, minute by minute

We didn't script a sales pitch — we recorded the actual product doing the four things it exists to do, back to back, on a real (demo) organization's data.

1. 0:00 — Connect a provider. Settings → Providers → paste an OpenAI Admin API key. That's it. The key is encrypted at rest with Fernet before it ever touches disk, and the first sync — 30 days of backfilled history — kicks off in the background the moment the credentials test clean.
2. 0:20 — The dashboard populates live. Cost, tokens, and request volume broken down by provider and model, updating as new data lands. This is the same TimescaleDB hypertable + continuous-aggregate setup that keeps a 30-day query under 50ms even past 100M events — you're watching production infrastructure, not a demo shortcut.
3. 0:45 — An anomaly gets flagged. The z-score baseline catches a spend spike against the organization's own 7-day history and raises it as an insight, the same way it would page a Slack channel in production.
4. 1:05 — A recommendation, with a number attached. The recommendation engine has already noticed that most of this workload is running on a model more expensive than the task needs, and it doesn't just say "consider a cheaper model" — it says switch to gpt-4o-mini and save $432/month, computed from this organization's actual 30-day token volume.

Why it's built this way

None of this is a wrapper around someone else's dashboard. The backend is FastAPI with SQLAlchemy 2 async end to end; the time-series layer is PostgreSQL 16 with TimescaleDB hypertables and continuous aggregates, chunked and compressed so a 30-day dashboard stays fast whether you have 10,000 events or 100 million. Every query is scoped by organization_id — multi-tenancy isn't a filter bolted on later, it's the first thing every route checks. Provider credentials never sit in plaintext: Fernet encryption at rest, decrypted only at the moment of the API call.

The anomaly detector and recommendation engine aren't black boxes either — they're plain z-score baselines and a substitution-rules engine that computes projected cost from your own token volume, not a generic benchmark. You can see exactly why a number showed up.

Try it on your own numbers

The $432/month in the video is one organization's real 30-day workload. Every account sees its own number, computed from its own usage — connecting your first provider takes about the same 30 seconds it took in the video.