Ilham Orujov

CEO and co-founder

Diablo AI

AI that evolves AI. When an AI system gets worse, Diablo finds which change caused it, and proves it.

diablo.pnoia.devThe live product
The two-minute demo.

Proof, not promises

  1. 40/40 → 0/40

    A recorded run in production, 9 Oct 2026: two changes shipped at once. Diablo found the prompt change that broke the system and cleared the innocent one.

  2. ≈ $0.20

    The model cost of that whole investigation, 123 calls.

  3. 97.3%

    On a seeded benchmark of 45,000 simulated regressions, it named the right cause 97.3% of the time when it named one.

  4. 1.1%

    False alarms when nothing had caused a change, against 81% for the usual "blame the biggest drop" approach (same simulated benchmark).

  5. 0

    Numbers the AI writes itself. Code computes every number, and a checker rejects any number the model tries to write.

  6. 293

    Unit tests. Its statistics match SciPy and statsmodels to within 5×10⁻⁵.

How it works

  1. You ask why your AI changed, in plain language.
  2. Diablo turns the question into competing hypotheses and designs one controlled experiment for each.
  3. Code runs your AI on the same cases in both arms and computes every number.
  4. You get a verdict per cause, with its effect size, confidence interval and evidence grade.

Next: connecting any AI system, then noticing a change, investigating it and verifying a fix on its own.

My part

CEO and co-founder. We are a founding team of two.

From the demo

This is Diablo, an AI that evolves AI.
This is Diablo, an AI that evolves AI.
Accuracy fell from 91 to 78 percent. We ask Diablo: why?
An update made the AI worse. Diablo is asked why.
Three hypotheses: the prompt, the index, the model.
Three competing hypotheses, one experiment each.
The AI reasons, the system measures. Diablo never invents a number.
The AI reasons. The system measures.

The example in the demo is simulated, and labelled so on screen.