AI for Scalable Tech Support
A platform architecture consulting project for a global software firm for demand forecasting using a configurable and highly scalable forecasting system for a significant cost optimization and NPI improvement program across channels and geographies.
AI for Scalable Tech Support
A platform architecture consulting project for global software firm for demand forecasting using a configurable and highly scalable forecasting system for a significant cost optimization and NPI improvement program across channels and geographies.

Objectives
- Improve forecasting accuracy from 65% to 95+% on a $700+ million support budget outlay (spent in L1/L2 customer support).
- Build an AI solution for real-time demand forecasting to optimize and transform workforce management.
- Build actionable insights for financial planning, resource planning, and channel investment plannings.
- Support forecasting leads to run and fine-tune forecast every cycle.
Challenges
- Data quality and availability; highly manual and broken data sources.
- 50+ channels for support incidents.
- Different forecasting algorithm across channels and geographies.
- Delay in forecasting availability to different support and corporate functions.
- Business value add and data science functions not in sync.


Solution


Implemented machine learning models and ensemble techniques to address seasonality, low volume behaviour, and small dataset issues for high quality forecast.

Unified data hub architected for curated data across global support tools using Azure Data Factory, Data Lake, and Azure SQL.

Daily, weekly, and monthly forecast per business need using statistical models and snapshots.

Technology
Azure Data Factory, Azure Databricks, Azure Data Lake, Synapse, Python, Spark, Lakehouse, Azure Purview, Message Hub, Azure Data Lake Storage Gen2, Azure Blob Storage, Event Hubs, scikit-learn, H2O
Anaplan for productivity modelling


Outcome

A 100% automated,
scalable engine to deliver forecast on demand.
Improved forecasting frequency from a three-month cycle to monthly.
97.6% forecasting accuracy delivered (from 65%).
120 locales supported; 40,000+ time series supported.
35% cost reduction; NPI improved by 45%.
97% automated anomalies engine.
99.7% platform availability.
