Audio intelligence for tele-sales

Audio in.Actions out.

Viveka listens to messy, real-world Hinglish insurance calls — 8 kHz, noisy, code-switched — and returns two machine-readable profiles per call. Every extracted fact carries a verbatim quote and a confidence level. Nothing is asserted without a receipt.

  viveka · live extract
Lakhs
of calls analysed
18
Profile field groups
~60
Fields extracted
4
Speaker roles diarized
0
Guesses made
01 — Diarization

Hears the whole floor.
Reports only your call.

Neighbouring agents read the same script into the same open-plan mic. Viveka decides what's real by who actually responded — not by who's loudest — and drops the bleed-through.

02 — Evidence-bound extraction

Click a fact. See the receipt.

Any field above low confidence must quote a verbatim substring of the transcript. Below that, the value is forced to not_disclosed and evidence is left empty. The model can't confidently assert what the customer never said.

Transcript · call 2026-05-27Hinglish · romanized
03 — Length-routed analysis

The right tool for the right call.

A 20-second voicemail and a 12-minute discovery call are different problems. Each transcript is routed to a length-specialized extractor.

04 — The dossier assembles

One call becomes a dossier.

From a single transcript, Viveka builds an 18-group profile — then a deterministic layer computes BANT, bands and the lead grade in code, so the same call always grades the same way.

Customer profileevidence-backed · confidence-scored
Lead quality — BANT, scored 0–3 each
BANT total 0/12 → lead grade
Deterministic. Computed by code, not the model.
The standout feature — contextual pain points

Not that they objected — why.

Most tools give you a flat label. Viveka gives you the reason, in the customer's own words, and how badly it matters. A dead label becomes a conversation the next agent can actually have.

05 — Product-family-aware compliance

The real violation, not the 96% that aren't.

"Guaranteed 35% return" is compliant on a non-par endowment and a violation on a market-linked ULIP. A naive keyword flagger fires on the word and mislabels most calls. Viveka classifies the product family first, then judges every claim in context.

96%
of calls a keyword flagger falsely marks as violations
0
genuine high-risk calls surfaced — not thousands of false positives
06 — The intelligence layer

Ten thousand calls become a pipeline.

Live campaign view — lakhs of calls
evidence-scored · mean score 84.3 / median 87

Where the pipeline stands

Every call auto-graded A–D. 43% are A-grade leads to call back today; 12% are C/D to deprioritize.

The skill cliff

Six coaching dimensions, averaged by grade. Watch where each collapses A → D.

What predicts a good call

Correlation of each signal with the overall call score.

Where agents lose the call

The intelligence isn't the score — it's knowing what to coach. Each step down grade is a different failure.

Agent leaderboard

Top performers by average score — with the signals that separate them.

07 — At scale

One card. Then the whole floor, self-sorting.

Under the hood

Two feeds. One structured output.

Vk
Self-hosted

Private inference endpoint

Standards-compliant API — streaming, strict JSON-schema and tool calling. Quantized for high single-node throughput.

Managed

Elastic managed fleet

Multi-key rotation and thread-pool parallelism. Built to grind an entire call center's daily volume, unattended.

Forced function calling Constrained JSON decoding Bracket-repair salvage Idempotent runners Deterministic scoring
Honest scope

One call → one profile. It doesn't yet stitch a customer's calls into one journey.

Not a SOP grader

It doesn't score calls against a specific company call-script.

Domain-tuned

Built for Indian life-insurance tele-sales. That focus is the source of the accuracy.

Request access

It doesn't just transcribe.
It understands.

18 field groups. ~60 fields. Zero guesses. Every fact comes with the quote that proves it.