Improve demand forecast accuracy by at least 10% for each SKU for each location at weekly level
Segmented SKUs to homogeneous clusters based on their sales patterns and used 18 machine learning algorithms to choose the best model for each SKU in an automated manner
The adjacent graph represents the improvement in forecast accuracy for each quadrant using our Automation model (shown in green)
Each SKU was classified into quadrant based on its
historical sales pattern
APO Model is the existing model used within SAP (shown in blue)
Automation Model was developed by us
Client needed to create an automated system which can read the email and understand the customer intent and subsequently allocate it to the right team within the CRM system
Used Natural Language Processing (NLP) to analyze customer emails and used unsupervised topic modeling to assign topics for each incoming emails. These intents were categorized into Level 1 and Level 2 intents to allocate to various teams
Python
Machine learning algorithms
LDA
Topic modeling
Understand Customer Satisfaction with Social Media Analytics
Problem:
The client was receiving a lot of comments on social media about its products and services. Their inhouse was not capable of finding each social post and reply to them in time
Objective:
Client needed to create an automated system to collect all tweets to start with and enable a mechanism to respond to them
Approach:
Developed a solution which connected to twitter API and collected all tweets that mentioned our client. All tweets were parsed and analyzed using NLP. Sentiment scores were used to classify tweets into various buckets. Customers with negative Sentiment were escalated for faster response and protect the brand online.
Benefits
Technologies:
Used R for collecting all tweets
Were able to capture all tweets related to the client by monitoring appropriate “hashtags” and “keywords”
Limited the tweets to particular geography and language indicator
English language grammar database was used for parsing, lemmatization
Stemming was used to narrow down the concepts
Sentiment scores were calculated using linguistic meaning in an automated manner
Each tweet was scored and stored for further processing
Tweets with high negative scores were escalated for faster processing