Business Objective
Our client is a leading multinational telecommunications conglomerate headquartered in Singapore.
The client was looking to leverage historical data to accelerate the process of submitting first cut quotations to large enterprise customers which took 15-27 days. The objective was to develop a price recommendation engine to generate costs faster and eliminate manual dependencies using ML/AI models in combination with the then-existing rules engine to pre-estimate the cost elements like Monthly Recurring Cost (MRC) and One-time Cost (OTC)
Challenges
- Not enough data points to build the model, and missing data in some cases
- Under predictions that lead to loss of leverage in the subsequent negotiations, thus leading to lower margins
- Difficult to capture the trends in data due to external factors like frequent cost changes, multiple negotiations, etc
- Data lacked some of the key features like on-net, off-net details (infrastructure data)
Solution MethodologyÂ
- Designed a custom loss function which accounts for the risk asymmetry by penalizing under-predictions more than over-predictions
- Derived additional features related to customer site location, customer firmographic attributes, partner-specific attributes, and historical pricing to customers from both internal & external sources to improve the accuracy
- Developed models to predict OTC and MRC with individual models for each country and cost type
- Developed real-time serving layer on Cloudera and integrated with the existing quotation platform
- Developed a model monitoring dashboard to track the model performance and input data drift
Business Impact
- Reduced Turn Around Time (TAT) for developing the first stage pricing from 15 days to 2 days
- Developed models for 7 countries to predict OTC (one-time cost) and MRC (monthly recurring cost) for multiple products with ~70% in range accuracy
- Developed dashboards to monitor the performance of the model predictions and refresh the models if performance degrades over time
