Questions about Kirelta
What is Kirelta?
Kirelta provides a monitoring service for production machine learning models that observes feature data sent to a model and issues verdicts on whether predictions remain trustworthy, degraded, or untrusted. It operates through an API with three steps—fitting a baseline on healthy data, assessing new batches, and branching on the verdict—without requiring access to the model itself or raw records. The service includes a live console, example decisions, and a free plan covering two models.
What are the pricing tiers and prediction limits?
The free plan covers two models and includes one million predictions per month, while the Pro plan supports ten models and twenty-five million predictions per month. The Business plan allows forty models and one hundred fifty million predictions per month, and the Enterprise plan offers custom models and predictions with unlimited users. All plans include verdicts with evidence, a calibrated drift alarm, and alert channels like email, Slack, webhooks, PagerDuty, and Datadog.
How does the API workflow function?
Kirelta uses a three-step API process: first, you fit a baseline by sending healthy data; second, you assess recent rows to receive a verdict and evidence; third, you decide to serve, flag, or hold based on the output. The system requires only the same feature vectors you already compute, needing no access to your model code or raw records. Each verdict includes plain fields showing which features carry the deviation and a recommended action.
What happens when I refit the baseline?
The baseline is the only ground truth the engine assumes, and refitting on drifted or biased data will make future verdicts reflect that new baseline instead of the original one. There is currently no versioning of baselines, so if you need to compare against an older one, you must keep the training rows yourself. Calling /fit again with different data simply replaces the trained baseline for that specific model.
How does the system handle prediction limits?
When you hit your monthly prediction limit, new calls to /assess are refused with a 429 response until the counter resets next month or you upgrade. The system fails closed with a clear error rather than silently dropping requests or degrading accuracy, ensuring you are aware of the limit. Billing is based on predictions processed per calendar month, not the number of API keys used.
How long is data kept and can I delete my account?
History retention varies by plan: Free retains data for 7 days, Pro for 30 days, Business for 90 days, and Enterprise offers effectively unlimited retention. You can delete your account and everything in it from Settings, which removes the account, models, keys, uploads, and decision history. An automated test proves the deletion leaves nothing behind except a receipt with the internal id and deletion time.
Is the free plan a trial?
The free plan has no trial clock or card requirement and is free indefinitely, not a 14-day countdown. If you outgrow its limits, upgrading is immediate and self-serve without any expiration. The Pro plan costs $99 per month or $82.50 per month billed yearly, while Business is $349 per month or $290.83 billed yearly.
What is the difference between DEGRADED and UNTRUSTED?
DEGRADED means the flagged rate is above twice the expected false-positive rate, but no alarm has fired yet, making it worth a look before escalation. UNTRUSTED means the drift alarm fired or more than 20% of the batch was flagged, requiring a block or escalation to a human. The recommended action for UNTRUSTED is to block or escalate, while DEGRADED is worth monitoring but not necessarily stopping service.
How does the decision journal and audit trail work?
The system includes a decision journal where a person approves or overrides each call with a reason, and the evidence is sealed by its hash and exports as a PDF. Every account and key action is chained by hashes and signed at checkpoints, allowing verification in the browser or offline with the published verifier. This creates a signed audit trail for every interaction.
How does the drift alarm and alerting system operate?
The system includes a calibrated drift alarm with a false-alarm rate bounded across the whole run, not per batch, and every response states it. Alerts are sent via email, Slack, webhooks, PagerDuty, and Datadog, with webhook channels offering a test you can send first. The evidence panel shows exactly which inputs moved and by how much, plus the raw evidence value against the alarm threshold.
What is the official website of Kirelta?
The official website of Kirelta is kirelta.com.
Does Kirelta have reviews on Trustburn?
Kirelta has no reviews on Trustburn yet. Anyone who has used it can write the first review on its Trustburn page.
AI-generated from reviews on this page. Verify anything important with the company directly.
News
No news yet. This is where the business shares its own updates and press.
Photos
Widgets
Get the widget for your site →About Kirelta
Kirelta provides a monitoring service for production machine learning models. It observes the feature data being sent to a model and issues verdicts on whether its predictions remain trustworthy, degraded, or untrusted, along with supporting evidence and a recommended action. The service operates through an API with three steps — fitting a baseline on healthy data, assessing new batches, and branching on the verdict — without requiring access to the model itself or to raw records. A live console and example decision are shown on the site, and a free plan covering two models is offered.
- Website
- kirelta.com