Inside Homayar: The Anatomy of a Multi-Agent Call Analysis Pipeline

When teams evaluate AI voice analytics, they usually ask about the model. The more useful question is about the pipeline — the sequence of steps between raw audio and a number a supervisor is willing to act on. This post walks through Homayar's pipeline stage by stage, including the design decisions that matter most for accuracy and auditability.

Stage 1: From audio to structured conversation

Everything starts with transcription, and telephony audio is hostile terrain: 8 kHz sample rates, compression artifacts, crosstalk, hold music, and accents the demo datasets never saw. Our speech-to-text stage produces more than a transcript — it produces a structured conversation: speaker-separated turns (diarization), word-level timestamps, and acoustic features such as silence gaps, talk-over events, and speaking-rate shifts that downstream agents use as signals.

Before anything leaves this stage, PII redaction runs. Card numbers, government IDs, and other sensitive spans are masked in both transcript and stored audio references. Downstream agents never see what they do not need.

Stage 2: Why specialized agents beat one big prompt

The naive architecture sends the whole transcript to one large model with one large instruction: score it, find issues, check compliance, summarize. It produces output that looks complete and measures poorly, because the model is averaging its attention across ten jobs.

Homayar instead runs a team of narrow agents, each with a single mandate and its own evaluation harness:

• The QA scoring agent applies your rubric — greeting, discovery, resolution, closing — and nothing else.
• The issue detection agent identifies what the customer actually contacted you about, including secondary issues raised mid-call.
• The compliance agent verifies required disclosures and prohibited statements against a versioned rule set.
• The sentiment and effort agent tracks emotional trajectory and customer-effort language across the call, not just an endpoint average.
• The insight agent extracts business signals: competitor mentions, product friction, process failures.
• The coaching agent converts the other agents' findings into specific, evidence-linked development points per agent.

The practical benefit is calibration. Because each agent does one job, we can measure its agreement with human experts on that job alone, tune it independently, and upgrade it without destabilizing the rest of the system. A monolithic prompt offers no such seam.

Stage 3: Orchestration and disagreement

Agents run in parallel against the same structured conversation, and an orchestration layer reconciles their outputs. Disagreement is treated as information. If the QA agent scores resolution highly but the issue agent marks the customer's primary issue as unresolved, the call is flagged for review rather than averaged into a dashboard. Every agent also emits a confidence value; results below threshold route to a human review queue instead of silently entering your metrics. We would rather show you a smaller number you can trust than a larger one you cannot.

Stage 4: Evidence binding

This is the rule that shapes everything else: no claim without a citation. Every score component, every compliance determination, every coaching point is bound to specific transcript spans with timestamps. When a supervisor sees "rushed discovery, 71/100," one click plays the forty seconds that produced it.

Evidence binding is what separates an analytics product from an opinion generator. It is also what makes the companion model real: an agent who disagrees with a score is not arguing with a black box — they are looking at the same evidence their supervisor is, and sometimes the evidence vindicates them. That has to be possible, by design.

Stage 5: Aggregation and trend detection

Per-call results roll up into per-agent, per-team, and per-driver views. A separate trend layer watches the aggregates for movement: a contact driver growing 30% week over week, a disclosure rate slipping on one queue, a new-hire cohort plateauing early. These are the outputs operations leaders actually run meetings on.

What the architecture buys you

Three properties fall out of this design. Auditability: every number traces to evidence, which matters enormously in regulated industries. Honest uncertainty: low-confidence results are surfaced as such instead of laundered into averages. Independent evolution: each agent improves on its own track, validated against its own human-agreement benchmarks, without regression risk to its neighbors.

None of this is exotic. It is the same discipline you would apply to any measurement system that people's performance reviews and your compliance posture depend on. The model matters. The pipeline around it is what earns trust.