From recording to receipts. Zero black box.

Every call streams its stage telemetry over the API — GET /api/v1/calls/:id shows where it is and what it found. This is that view.

1. 01 · ingest

   Ingest — every recording, one pipeline

   Push recorded calls over the REST API as they close, drop a day's batch, or wire your telephony to do it for you. Any common audio format — and nothing gets sampled, because everything gets analyzed.

   call_intake — 2026-07-10

   1,847 recorded calls

   api_upload 1,204 · batch 512 · connector 131

   formats mp3 · wav · m4a — fair use 8 min per billed call

   webhook registered: job.succeeded → your endpoint

2. 02 · transcribe

   Transcribe — diarized, word-timed

   Our transcription engine separates speakers and timestamps every word, across 90+ languages — with best-in-class accuracy on Saudi and Egyptian Arabic. Dialects are transcribed as spoken, never flattened through translation.

   transcript.json — 9f2c51b8-4a07

   lang ar-SA · speakers 2 · words 1,082 timestamped

   00:41 العميل: انخصم مني نفس المبلغ مرتين على فاتورة يونيو، وش الحل؟

   00:52 الموظف: حياك الله، أعتذر لك عن الإزعاج — أتحقق من الفاتورة الآن

   Saudi dialect, transcribed natively — not translated

3. 03 · redact

   Redact — PII out before analysis

   Names, phone numbers, cards, and addresses are masked in the transcript before any analysis model reads it. Security-first by construction — the model grades the conversation, not the customer.

   redacted_transcript — what the model sees

   name سلطان الحربي

   phone +966 55 481 2907

   card 4532 88•• •••• 1096

   dialogue kept · timing kept · meaning kept

   masked before the analysis model reads a word

4. 04 · analyze

   Analyze — the call becomes signal

   Frontier AI reads the redacted transcript and returns the summary, sentiment overall and per speaker, topics, pain points and service gaps, churn and escalation risk — in the language of the call.

   analysis.json — 9f2c51b8-4a07

   sentiment overall −0.34 · customer −0.61 · agent +0.22

   csat_predicted 2.5 / 5 · churn_risk 0.71 · escalation medium

   topics billing · double_charge · refund_timeline

   pain_point"refund timeline unclear" — 214 calls this week

   summary + findings returned in the call's language

5. 05 · score

   Score — QA with receipts

   A weighted scorecard — six criteria by default, fully yours to edit — grades every call, and every deduction cites the exact transcript line it came from. Low-confidence scores are flagged needs_review instead of pretending.

   qa_scorecard — 9f2c51b8-4a07

   greeting & verificationw684

   discoveryw770

   resolution ownershipw955

   empathy & tonew776

   compliancew852

   wrap-up & next stepsw582

   weighted total 67.8 · needs_review false · every deduction cites its line

6. 06 · coach

   Coach — said → better → why

   Every scored call yields concrete rewrites: what the agent said, the stronger line, and why it works. Team leads get coaching they can paste into a 1:1, not a dashboard to decode.

   coaching — agent_207 · call in en

   # said

   "That's our policy — there's nothing I can do."

   # better

   "I get why that's frustrating — here's what I can do right now."

   # why

   keeps ownership; policy-deflection is the top driver of 1–2 star CSAT in this queue