Every score points at the turn it came from
Merivex is a quality-intelligence layer, not a broad CX suite. It does one job — evaluate every connected conversation and turn the results into coaching — and it does it with an evidence trail the rest skip. Here is the whole pipeline.
Input
- Connected transcripts
- Your rubric
- Reviewer scores (optional)
Outcome
- Evidence-cited evaluations
- Patterns backed by multiple interactions
- Per-person coaching plans
Connect, read-only
Every connector issues read queries and schema lookups. No code path writes, updates or deletes anything in a system you connect.
A read-only database account is sufficient, and is the one Merivex would rather you used. Outbound connections verify TLS by default; a private certificate authority is configured explicitly, never by disabling verification.
A read-only database account is sufficient, and is the one Merivex would rather you used. Outbound connections verify TLS by default; a private certificate authority is configured explicitly, never by disabling verification.
Normalise and classify
Incoming interactions are validated and normalised into one shape, then classified: intent, the emotional movement through the conversation, and the turns where it changed.
Classification runs on every ingested interaction, not a sample. The intent taxonomy is maintained rather than left to drift, so “billing” means the same thing this month as last.
Classification runs on every ingested interaction, not a sample. The intent taxonomy is maintained rather than left to drift, so “billing” means the same thing this month as last.
Evaluate with citations
Each interaction is scored against your rubric. Every score carries the specific turn it was read from, so the score is arguable on facts rather than defended by seniority.
Compliance and policy-risk signals are evaluated in the same pass. The output is not a number in isolation — it is a number with the sentence it came from attached.
Evidence, not just a number
Agent4That's just how the refund timing works, there's nothing I can do about it.
Check the check
A separate verifier agent re-reads the draft evaluation and flags any claim the cited evidence does not actually support, then the evaluation is corrected before anyone sees it.
This is the step that lets Merivex say no conclusion is published without the evidence behind it. The verifier holds no special authority — it operates inside the same tenant and permission model as every other agent.
This is the step that lets Merivex say no conclusion is published without the evidence behind it. The verifier holds no special authority — it operates inside the same tenant and permission model as every other agent.
Patterns with a threshold
A behaviour is only recorded as a pattern once several interactions agree on it. One interaction is an anecdote; the platform will not promote it.
Each pattern keeps its evidence — the interactions and turns that support it — so a manager can open it and read the actual conversations, not just a count.
- Dismissive close14
- Missing verification step9
- Unclear refund timeline6
- Strong de-escalation21
Recorded only when multiple interactions support them.
Coach a named person
Evaluation evidence becomes a per-person coaching plan: what to change, which interactions show it, and a focus for the cycle.
Findings land on a specific workforce agent as something specific to do, with the interactions cited. A coaching note with no evidence attached is not something the platform produces.
Closing turns drop empathy scoring while resolution stays high. Three interactions this week end without acknowledging the customer's frustration.
Focus for this cycle
- Acknowledge before closing
- Offer the next step explicitly
- Confirm understanding
evidence · 3 interactions cited
Calibrate against your reviewers
Import the scores your human reviewers already gave, and Merivex compares its evaluations against them so you can see where the model and your team diverge.
Calibration is how you decide whether to trust the automated score for a given rubric dimension before you rely on it — not a claim that the two always agree.
Calibration is how you decide whether to trust the automated score for a given rubric dimension before you rely on it — not a claim that the two always agree.
What this is not
The honest boundary.
Merivex evaluates conversations and coaches the workforce. It is deliberately not the rest of the stack.
It does not answer customers or take actions in your systems. Read-only, by design.
It does not replace your reviewers. Calibration exists so you can decide dimension by dimension how far to trust the automated score.
It does not claim certifications it does not hold. The trust centre lists every gap.
It does not surface a pattern from a single interaction, however striking that one is.
See it run before you talk to anyone
The interactive demo uses seeded data and needs no sign-up. It is the same pipeline you have just read about, at small scale.