EveryCall: on-device QA that grades 100% of BPO calls for less than the price of sampling 3%

The problem

Indian BPOs manually review 2-5% of calls; one QA analyst gets through 20-30 calls a day against tens of thousands of daily calls, so compliance violations hide in the unreviewed 95% and international clients are demanding tighter oversight. Cloud auto-QA costs $50-180K a year and ships recordings to third-party clouds, a non-starter for cost-sensitive BPOs and their data-residency-paranoid clients.

The idea

Why now

Tether's TranscribePsy, open-source 254M on-device speech recognition, launched Oct 15 and collapses the transcription cost that made 100% coverage expensive; Jev 1.14 (Oct 15) brings the judgment layer for calibrated rubric scoring; and July-August 2026 industry pieces show BPOs actively migrating from 2% sampling to 100% reviews, so the budget line is opening now.

What it combines

tether-qvac-transcribepsy-open-source-254m-on-device-speech-recognition (on-device ASR) + jev-1-14-0-typesafe-s-judgment-layer-update (calibrated rubric judgments) + decision-3-0-open-multimodal-decision-models-0-6b-27b (small local decision models for escalation triage). The mix matters because transcription alone does not grade calls and scoring alone is unaffordable at cloud rates; on-device ASR plus calibrated local judgments is what makes 100% coverage cost less than 3% sampling.

MVP

Weekend MVP: ingest one day of call recordings from a single dialer API, transcribe on-device with TranscribePsy, score against a 5-question rubric with the Jev judgment layer, and output a ranked list of the 20 calls the QA head should review. Deliberately skip: real-time scoring, the coaching dashboard, and multi-dialer integrations.

How it makes money

The moat

Observe.AI, Level AI and CallMiner are cloud per-seat SaaS businesses selling $60K-180K/year contracts; a $5/agent on-device product cannibalizes their ARR, and matching it means rebuilding their cloud STT-plus-LLM pipelines for offline edge deployment, an architecture inversion rather than a feature. They cannot match the price without destroying their own business model.

Distribution

B2B2C through the dialer and CCaaS platforms BPOs already run (AirCall, Dialpad, CallHippo app marketplaces, a channel Enthu.AI has already proven) plus BPO consultancies that sell QA transformation. Land with one 200-seat floor, expand floor by floor.

Why it wins

Observe.AI, Level AI, CallMiner, Verint and Enthu.AI all score calls in their cloud at per-seat SaaS prices with multi-month implementations. EveryCall runs fully on the BPO's hardware at roughly a tenth of the price, recordings never leave the building, and it adopts the client's existing scorecard instead of a 3-6 month calibration project.

Risks

Biggest risk is transcription accuracy on accented, noisy audio: if word error rate is bad, every score is garbage. De-risk by measuring WER on the customer's real recordings in week one and routing low-confidence calls to humans by design.

Build it with

• Tether QVAC TranscribePsy: open-source 254M on-device speech recognitionzero-marginal-cost transcription of every call on the BPO's own hardware
• Jev 1.14.0 — TypeSafe's judgment layer updatecalibrated rubric scoring with confidence for escalation triage
• Decision 3.0: open multimodal decision models (0.6B-27B)small local decision models for cheap escalation triage

Repo to start from

everycall-open: the on-device ASR plus rubric-scoring pipeline and the scorecard schema; the hosted multi-tenant service and dialer connectors stay closed.

Evidence

• Google Cloud Conversational Insights pricing (Quality AI $50/agent/mo)
• Observe.AI pricing guide 2026 (CloudTalk)
• We Only Manually Review 3-5% of Calls (Pearl, Aug 2026)
• Why Are BPOs Moving from 2% Sample QA to 100% Interaction Reviews?
• Why QA Is Important to Every Call Center (Enthu.AI)