Just Grade Metrics — AI Call Quality Assurance

Contact centres record every call, then score one or two per cent of them by hand.

The sample is too small to be representative, scoring drifts between reviewers, and when an agent disputes a mark there is rarely anything to point at.

HoverTap a card to read what it prevents.

• No evidence, no score

  No evidence, no score

  A QA score a supervisor cannot trace to something someone actually said is worthless in a dispute. The check holds back roughly a third of real uploads — that is it working.

• Uncertainty shrinks the score

  Uncertainty shrinks the score

  When speaker attribution on a mono recording is not confident, criteria that depend on who spoke are marked not applicable rather than guessed.

• PostgreSQL enforces isolation

  PostgreSQL enforces isolation

  Row-level security on 31 tables, so an agent querying a colleague’s call is refused by the database, not filtered by the interface — tested by attempting cross-tenant reads.

• The model’s key is never on the web server

  The model’s key is never on the web server

  All model access lives in edge functions, so compromising the web server exposes no spend-capable credential.

• A published scorecard is frozen, prompt included

  A published scorecard is frozen, prompt included

  A score disputed months later can be answered with the exact instructions that were in force at the time.

• A human override keeps the original

  A human override keeps the original

  Both scores are retained. The gap between them is the only measure of whether the model can be trusted.

• Just Grade Metrics is a client’s product, built under contract from the first commit. The engineering is mine; the product is not.
• About a third of uploads go to a person rather than straight to a score. Budget for reviewers, not full automation on day one.
• Mono recordings are weaker than dual-channel: speaker attribution is inferred, and criteria that depend on who spoke may come back not applicable.