RotorPricingSign inGet started

An SQL database for AI-powered workflows

Build workflows and apps that make their own decisions. Rotor lets your agent answer questions like “Why was this decision made?” and “Would a different model have decided better?”

Unlock the potential of decision models

In the next generation of software, models will be making decisions as applications run. Their answers become inputs to other decisions, change data, and lead to actions. Reading the code no longer tells you why something happened. Asking the model again may give you a different answer.

Your agent needs to see the decisions actually made, follow their connections in both directions, and run them again:

What led to this decision?

  • The data as it was at the time.
  • The code and instructions used.
  • Earlier decisions it relied on.
  • What the model saw and what it answered.

What followed from this decision?

  • The data changed using its answer.
  • Later decisions that relied on it.
  • Actions taken as a result, through many steps.

Would a different model have decided better?

  • Change the model, the instructions or the code.
  • Replay it on the data exactly as it was.
  • Compare with the old answer and what happened.
  • Do it for one decision or a year of them.

When a mistake comes to light, an explanation is only the beginning. Your agent needs to try a different approach, inspect what changed since, and record a carefully chosen correction.

Rotor keeps recorded work in one database.

  • Read your data exactly as an earlier decision saw it.
  • Query dependencies your application explicitly records.
  • Record revisions alongside earlier versions.

Example: Test a change on a year of past decisions

A fictional hospital app, with illustrative figures.

  1. A hospital app checks every patient’s chart every hour and decides whether to raise the alarm for a heart attack.

    Your app, every hour, for every patient

    Watch for heart attacks

    1. Read the patient’s chart: vital signs, lab results and nurses’ notes.
    2. AI summarizes the latest notes.
    3. Check the hospital’s heart attack guideline.
    4. Decide: no alert, watch, alert the nurse, or call the doctor.
    5. If alerting, AI drafts the message.
  2. You rebuild it. It now reads medication records, tracks six-hour trends and follows the hospital’s new guideline.

    Your app, rebuilt

    Watch for heart attacks

    1. Read the patient’s chart: vital signs, lab results, nurses’ notes and medication records.
    2. AI summarizes the latest notes.
    3. Work out six-hour trends in heart rate, blood pressure and blood tests.
    4. Check the hospital’s new heart attack guideline.
    5. Decide: no alert, watch, alert the nurse, or call the doctor.
    6. If alerting, AI drafts the message, with the trends behind it.
  3. Would it have been better? In Rotor, testing it on every hourly check of the past year takes three steps, each check reading the chart exactly as it was at that hour.

    1. Pick every hourly check from the past year.

      Show the code
      SELECT patient_id, checked_at,
             rotor_commit
      FROM checks
      WHERE checked_at >= '2025-11-09T00:00:00Z';
    2. Run the rebuilt app on each one, with the chart exactly as it was.

      A blood test result that arrived at 5 am stays out of the 3 am check, as it did for the app.

      Show the code
      for check in past_year:
          chart = db.at(check.rotor_commit)
          alert = watch_v2(check.patient_id, chart)
          save_replay(check, alert)
    3. Compare both versions with what actually happened.

      Show the code
      SELECT o.patient_id, o.heart_attack_at,
             r.v1_alert_at, r.v2_alert_at
      FROM outcomes o
      JOIN replays r
        ON r.patient_id = o.patient_id;

    Elsewhere, history has to be set up in advance and kept working as the app changes, and it’s rarely exact.

  4. For 61 of the 412 patients who had a heart attack, the rebuilt app would have raised the alarm earlier, by 2.3 hours on average, with 30% fewer false alarms.

    61 of 412

    patients who had a heart attack, flagged earlier

    2.3 h

    earlier, on average

    30%

    fewer false alarms

    Across 38,000 admissions over the past year, compared with what actually happened to each patient.

How Rotor compares

One AI app’s first year: a billion model calls and every change to its records.

PostgresPostgres + DuckLakeRotor
Day to day
Writing dataFastFastFast
Everyday queriesFastFastFast
Big reportsSlowFastFast
Explain a past decisionIf you build itIf you build itBuilt in
Keep every changeGaps go unnoticedGaps go unnoticedComplete
Experiment on an instant copy of the pastNot availableNot availableWhole history
Change your tablesSafeRiskySafe
Systems to run131
Estimated cost
Data written in a year10 TB10 TB10 TB
Stored after compression5.5 TB0.9 TB + 1.8 TB2.0 TB
ComputePay 24/7Pay 24/7Pay for what you use beyond the plan
Monthly bill, busy month$1,180$792$471
Monthly bill, quiet month$1,180$725$156

Estimates for one example app at each service’s list prices, November 2026. Your costs depend on your data and use.

Pricing

Billing

$19.99 /month

Billed $239.88 yearly

$29.99 /month

Billed monthly

Start your free trial

7 days free. No card needed.

Included every month

100 GB Rotor compresses your data, so you can often store over 500 GB of raw data within the 100 GB allowance.

  • CPU50 CPU-hours
  • RAM250 GiB-hours

Beyond the allowance

Storage $39.99 per TB / month

  • CPU$1/CPU-hour
  • RAM$0.05/GiB-hour
  • Billed by the millisecond, only when used.
  • Outbound transfer$9/TB
  • Incoming transferFree