Case Studies

Case Studies

Selected examples of technology systems we have designed, built and deployed. Client names are kept confidential unless permission has been granted.

01
Telecommunications

Network Anomaly Detection for a Leading Telecom Provider

The Challenge

A large telecom operator was experiencing network degradation events that were only detected after customer complaints, resulting in SLA breaches and reactive maintenance costs. The existing monitoring system generated thousands of alerts daily with no intelligent prioritisation.

Environment

Multi-vendor network infrastructure generating over 50 million telemetry events per day across core, transport and access network layers. Data was siloed across multiple OSS/BSS systems with no unified view.

Our Solution

We designed and deployed a real-time anomaly detection platform that ingests network telemetry via Apache Kafka, applies ML-based anomaly scoring and surfaces prioritised alerts to network operations teams through an operational dashboard.

Approach

Unsupervised anomaly detection using Isolation Forest and LSTM-based time-series models trained on 18 months of historical telemetry. Models were deployed as containerised inference services with automated retraining pipelines.

Architecture

Kafka-based streaming ingestion, Spark Structured Streaming for feature computation, ML inference layer on Kubernetes, PostgreSQL for alert storage, React-based NOC dashboard.

Outcomes

  • Reduced mean time to detect (MTTD) network anomalies by approximately 70%
  • Decreased false-positive alert volume by over 60% through ML-based filtering
  • Enabled proactive maintenance scheduling before customer-impacting events

Technologies

Apache KafkaApache SparkPythonScikit-learnTensorFlowKubernetesPostgreSQLReact
02
Enterprise Technology

AI-Powered Knowledge Assistant for an Enterprise Technology Company

The Challenge

A technology company with a large internal knowledge base — spanning product documentation, support articles, engineering runbooks and policy documents — was struggling with knowledge discovery. Employees spent significant time searching for information across multiple disconnected systems.

Environment

Approximately 40,000 documents across SharePoint, Confluence, internal wikis and PDF repositories. No unified search or question-answering capability. Documents updated frequently with no version tracking.

Our Solution

We built an enterprise knowledge assistant using a retrieval-augmented generation (RAG) architecture, allowing employees to ask natural language questions and receive accurate, cited answers drawn from the company's internal documentation.

Approach

Document ingestion pipeline with chunking, embedding generation using OpenAI embeddings, vector storage in pgvector, retrieval and re-ranking pipeline, GPT-4-based answer generation with source citation. Role-based access control ensures users only retrieve documents they are authorised to access.

Architecture

FastAPI backend, pgvector for semantic search, LangChain for RAG orchestration, React frontend, Azure AD integration for authentication, deployed on Azure Kubernetes Service.

Outcomes

  • Reduced average time to find internal information from 15 minutes to under 2 minutes
  • Achieved over 85% answer accuracy on internal evaluation benchmark
  • Adopted by over 200 employees within the first month of deployment

Technologies

PythonLangChainOpenAI APIpgvectorFastAPIReactAzure Kubernetes ServiceAzure AD
03
Retail

Customer Churn Prediction for a Retail Business

The Challenge

A retail business with a loyalty programme was experiencing high customer churn without early warning signals. Marketing campaigns were applied uniformly rather than targeted at customers showing early signs of disengagement, resulting in high campaign costs and low retention rates.

Environment

Three years of transactional data covering approximately 500,000 loyalty programme members. Data held in a legacy SQL database with limited analytical capability. No existing ML infrastructure.

Our Solution

We built an end-to-end churn prediction system — from data pipeline and feature engineering through to model deployment and a campaign targeting interface — enabling the marketing team to identify at-risk customers 30 days before predicted churn.

Approach

Gradient boosting classifier (XGBoost) trained on RFM features, purchase category patterns and engagement signals. Model deployed as a batch scoring pipeline running weekly, with outputs surfaced in a campaign management dashboard.

Architecture

dbt-based feature pipeline on Snowflake, MLflow for experiment tracking and model registry, weekly batch scoring on Airflow, results written to PostgreSQL, React dashboard for marketing team.

Outcomes

  • Identified at-risk customers with approximately 78% precision at 30-day prediction horizon
  • Enabled targeted retention campaigns with 40% lower cost per retained customer
  • Reduced overall churn rate in the loyalty programme within two quarters of deployment

Technologies

PythonXGBoostdbtSnowflakeApache AirflowMLflowPostgreSQLReact

Discuss a Similar Challenge

If you are facing a technology problem similar to those described here, we would be glad to discuss how we might approach it.