CLIO

CLIO ingests every call and every document. Calls get transcribed and diarized; documents get parsed. Both flow through an LLM grader running against a versioned rubric, with scores landing in ops dashboards within minutes. No sampling. No human grader bottleneck.

The rubric is the artifact. The model behind it is replaceable; the rubric is the spec.

Call and document QA traditionally means a human samples 5% of the volume, grades by hand, and hopes the sample is representative. Compliance becomes theatre. The other 95% goes ungraded, ungoverned, and unseen — and the worst calls and the riskiest documents are the ones nobody reads.

CLIO grades all of them.

The rubric, not the model, is the source of truth. A graded call is only as trustworthy as the rubric it was graded against. So the rubric is the artifact ops owns, versions, and edits. When ops disagrees with a score, they edit the rubric — not retrain the model. The model is plumbing; the rubric is the product.

• Rubric items, weights, and scoring scales — versioned, every change auditable
• Score thresholds — what counts as pass, fail, or escalate
• Coaching tags — what flags attach automatically to surfaced calls
• Routing — which scores go to which dashboards, queues, and owners
• Sampling overrides — pin specific calls into the regression set
• Coverage rules — which call types get scored on which rubric axes

1. Edit the rubric in the console.
2. Re-run the grader against a frozen sample of past calls.
3. Diff the new rubric scores against incumbent — see which calls get re-classified.
4. Promote when the rubric matches senior-grader consensus better than incumbent.
5. Every call from then on uses the new rubric. Rollback is one click; every version is auditable.

• Speech diarization
• Document parsing
• LLM-as-judge grader
• Rubric-versioned eval
• Realtime dashboards