Don't just detect drift.
Decide.
Kirelta watches your models, and the moment one drifts it tells you what broke, what it's costing, and exactly what to do — with the confidence behind every call. No labels. No model access.
Most tools stop at the alarm. The hard part lands on you.
"Drift detected" is where the work begins, not ends — why did it happen, how bad is it, and what do you do right now? Kirelta closes that gap: it turns a raw statistical signal into a decision you can act on in one screen.
Calibrated detection
Conformal p-values with coverage guarantees, tested by the KS statistic — not a threshold you re-tune forever. Fully unsupervised: no labels, no model access.
Decisions, not dashboards
Every alarm becomes a case: the cause, the estimated cost, and ranked actions with the confidence and risk behind each one. You decide in seconds.
Drops into your stack
A REST API and native /metrics for Prometheus. Sits in the request path at 33ms p50. Managed cloud or self-host the same package.
Three steps from raw data to a decision.
A distribution-free pipeline built from classical pieces — composed into one number, then one recommendation.
Learn what normal looks like
Kirelta fits a multivariate model of your healthy data — its shape, spread, and correlations. No labels required.
Turn surprise into a p-value
Each batch gets a conformal p-value and a sequential martingale — calibrated evidence that accumulates until it crosses.
Get the cause, cost, and action
Kirelta decomposes the drift, estimates the impact, and ranks the recommended actions by confidence and risk.
The difference is what happens after the alarm.
Hardened before it ever shipped.
Hashed credentials
Passwords PBKDF2-salted; API and email tokens stored only as hashes. A leak exposes nothing usable.
Multi-tenant isolation
Per-model boundaries, rate limiting on every sensitive route, and a full audit log of who changed what.
Your data stays put
Self-host the same package inside your perimeter — SQLite or Postgres, your infrastructure, your keys.
Honest by design
When a calibration guarantee doesn't hold on a stream, Kirelta says so — never a false green light.
Start free. Pay when it's load-bearing.
- 1 monitored model
- Calibrated drift detection
- Decision cards
- Community support
- Unlimited models
- Ranked actions + cost estimates
- Prometheus + exportable reports
- Priority support
- Docker, Compose, or Kubernetes
- SQLite or Postgres
- Your data never leaves
- Same API, same metrics
The questions engineers ask first.
Do I need labels or access to my model?
No. Kirelta is fully unsupervised — it reads the behavior your models already emit (inputs, scores, embeddings). Nothing to annotate, no model internals required.
Will it slow down my requests?
It's built to sit in the request path: 33ms p50, 88ms p99 under load, with thread-safe atomic model swaps. Or call it asynchronously in batches — your choice.
Can I run it inside my own infrastructure?
Yes. The self-host tier is the identical package — Docker, Compose, or Kubernetes, backed by SQLite or Postgres. Your data never leaves your perimeter.
How is this different from Datadog or Arize?
Those show you the signal. Kirelta closes the loop: it tells you the likely cause, the estimated business cost, and the ranked action to take — calibrated, and without requiring labels.
Turn your next model alarm into a decision.
Point Kirelta at a baseline, send your batches, and get a calibrated verdict — and the action behind it. The free tier is enough to see it work.