# A national mobile operator in Nepal cuts churn and SIM-box fraud with in-country analytics

> A national mobile operator in Nepal uses DataNature in its own private cloud to predict churn, detect SIM-box and subscription fraud, and optimize its network. Subscriber data stays in-country, inside the operator’s own environment.

Source: https://datanature.ai/customers/nepal-mobile-operator-churn-fraud
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

- Customer: A national mobile operator in Nepal
- Sector: telecom
- Country: Nepal
- Hosted on: Telecom private cloud, in Nepal
- Products: DataNature, AI360 Studio

## The challenge

Call records and network events arrived in huge volumes. Fraud was found weeks late. Churn models ran on stale extracts. Subscriber data had to stay inside the operator’s own environment.

## The solution

DataNature ran inside the operator’s private cloud. StreamHouse streamed call records and network events. AI360 Studio scored churn and fraud risk and synced next-best offers to campaign tools.

## The results

- Fraud patterns surfaced while they are happening
- Retention teams act on fresh churn scores
- Subscriber data kept in the operator’s own cloud

(Anonymized at the customer's request.)

## What was the operator trying to fix?

Fraud and churn were spotted too late to act. Analysts waited days for extracts. And subscriber data could not leave the operator's own cloud.

## What did it deploy?

DataNature in the operator's private cloud, with StreamHouse for real-time events, the Lakehouse for history, and AI360 Studio for churn, fraud and next-best-offer models.

## What changed?

Fraud patterns surface while they are happening, and retention teams act on fresh churn scores.

## How does StreamHouse power fraud detection here?

The architecture puts fraud on the real-time hot path and churn on the governed batch path, over the same tables in the operator's private cloud:

1. **StreamHouse** streams call detail records and network events through managed Kafka, keyed by subscriber, into Bronze Iceberg tables.
2. **In-flight detection** flags fraud patterns as traffic happens: SIM-box behaviour (many short outbound calls, few inbound, static location), call bursts, sudden roaming spikes.
3. **Alerts** reach the revenue assurance team in minutes instead of weeks, with the triggering events attached.
4. **Batch churn models** run on Spark over Gold subscriber history: usage trends, complaints, network experience, recharge patterns. AI360 Studio scores churn and syncs next-best offers to campaign tools with reverse ETL.
5. **Every event, score and decision** is written back to Iceberg, so models retrain on outcomes and investigators can replay any alert.

Subscriber identifiers are masked for analysts, and the whole stack, control plane included, runs inside the operator's environment.

## What can be added next on the same platform?

Churn scores and live usage data on the operator's platform can drive personalized Nepali voice calls: recharge reminders before a balance runs out, plan upgrades matched to real usage, and save calls for subscribers at risk, delivered through a Nepali voice platform such as TingTing with consent and contact limits enforced in Atlas. See [Nepali AI voice calls for banks and telcos](/blog/nepali-ai-voice-calls-banks-telecom-ai360-tingting).
