The problem
Organizations increasingly let AI systems score things that matter — sales calls, support interactions, student work, engineering output. Most of these systems are black boxes: a number comes out, and nobody can say why. That is unacceptable anywhere a score affects a person. If you cannot explain a score, you cannot appeal it, audit it, or improve it — you can only obey it.
The approach: a three-layer pipeline
Mizaan is Arabic and Urdu for balance — the scale of justice. The system is designed so that no single model's judgment is ever the final word:
- Layer 1 — Primary scorer. A scoring agent evaluates the input against explicit, human-readable criteria and produces a score with its reasoning attached.
- Layer 2 — Auditor. An independent agent reviews the primary score and its reasoning, challenging inconsistencies and flagging weak justifications before anything is reported.
- Layer 3 — Human validation. People stay in the loop by design. Disputed or low-confidence scores are routed to a human, and human corrections feed back into the system.
That feedback loop is powered by retrieval-augmented learning: validated decisions become reference material the scoring agents consult on future cases, so the system's judgment converges toward its human overseers' — visibly, case by case, instead of through opaque retraining.
Where it applies
The pipeline is domain-agnostic. It has been designed to work across sales, support, education, operations, and engineering — anywhere performance is scored and the score has consequences.
Status
Live at mizaan.mannatai.com and in active development, evaluated across two contrasting domains over three independent runs for the published study. Current focus: hardening the auditor layer and expanding the evaluation criteria library across domains.