Case Studies How We Work Blog About Book a call
AI Services & IT Outsourcing · Enterprise Digital Teams · NDA NDA

On-premise voice AI for a regional telecom operator

A regional telecom operator (12M+ subscribers, NDA) modernised voice self-service and call analytics on-premise: Parakeet ASR on 100% of recordings, a conversational agent for top subscriber intents, SIP/PBX integration and CRM webhooks — without cloud speech APIs or replacing core BSS/CRM.

Discuss a similar project All case studies

Service

AI Services & IT Outsourcing

Industry

Enterprise Digital Teams · NDA

Team

1 PM · 2 AI engineers · 2 Backend · 1 DevOps

Technologies

Parakeet ASR · LLM dialogue engine · SIP/VoIP · Python · Kafka · PostgreSQL · Redis · Kubernetes · Prometheus/Grafana

Client context

A mid-size regional telecom operator serving 12M+ mobile subscribers ran customer contact through a legacy IVR and a distributed call-centre network. Peak load exceeded 12,000 inbound calls per day across sales, billing, and technical support. Recordings were stored on the operator's PBX with no systematic transcription or quality control. After-hours and overflow traffic was lost or queued for minutes. Regulatory requirements mandated that subscriber voice data remain inside the operator's perimeter — cloud ASR APIs were not an option.

Challenge

The operator needed to modernise voice self-service and call analytics without replacing core BSS/CRM or moving data off-premise. Roughly 35% of routine enquiries (balance, tariff change, SIM status) still required a live agent. QA teams manually sampled under 2% of calls with no structured scoring or alerting. Any voice AI had to answer within 2 seconds and survive node failure without downtime. Results had to flow into the existing CRM and ticketing stack via API, not a separate silo — with a production pilot on 2 contact-centre sites within one quarter.

What we did

  • Deployed an on-premise voice AI platform on a 2-node HA cluster inside the operator's data centre.
  • Integrated Parakeet ASR for real-time transcription, diarisation, and post-call analytics on 100% of recorded interactions.
  • Built a conversational voice agent for top-15 subscriber intents (balance, top-up, tariff info, SIM block/unblock, escalation to agent).
  • Connected the platform to the operator's SIP trunk and PBX for inbound routing, overflow, and after-hours handling.
  • Delivered an analysis pipeline: transcript → LLM/rule-based scoring → checklist → webhook to CRM and supervisor alerts.
  • Implemented monitoring, failover, and staged rollout across 2 pilot sites before network-wide scale.

Process

  1. Discovery & audit — mapped call flows, intent distribution, PBX/SIP topology, CRM API, and data-residency constraints.
  2. Architecture & HA design — 2-node Kubernetes cluster, Kafka event bus, automated failover testing.
  3. ASR pipeline — Parakeet deployment, tuning on operator call corpus, diarisation and latency optimisation.
  4. Voice agent build — dialogue flows, intent/slot NLU, TTS integration, seamless handoff to live agents.
  5. CRM & alerting integration — REST webhooks, call summary, score, and escalation flags into existing ticketing.
  6. Pilot (2 sites) — shadow mode → limited traffic → full production on pilot sites.
  7. QA & observability — dashboards, SLA monitoring, weekly tuning loop with contact-centre supervisors.

Result and impact

Within 12 weeks of go-live on pilot sites, the operator cut routine agent load, improved call visibility across 100% of recorded interactions, and met data-sovereignty requirements — all without external speech APIs or replacing core BSS/CRM.

100% of recorded calls transcribed and scored

This is an NDA-protected engagement. Client name and identifying details are withheld; industry, region, challenge, solution and outcome are shared in an approved form.

Screenshot 1