Aman Kumar Garg

Designed and implemented data-driven marketing campaigns through Digital Notification channel

Personal Loan Model

• Built propensity model using spending data, events data etc to Identify customers who have higher conversion of taking personal loan
• Achieved the ROC value of 0.82 on Out of time dataset using LightGBM
• 414k+ audience reached with acceptance rate of 4.6% and 10k customers took loan through campaign annually

Balance Growth and Retention Model

• Developed a predictive model to identify potential depositors, enhancing deposit growth and customer retention for the bank.
• Identify key indicators such as transaction patterns, income levels, tenure and Credit card limit value to score customers on their likelihood to increase deposits.
• Targeted around >500k high potential customers through incentive driven campaign and increase in deposits to $50M from the campaign

Pricing Engine

• Built the dynamic pricing system for buying a used car based on car features, inspection report and demand-supply in the market
• Trained the regression model using Neural Network architecture and predicted the depreciation of a car from its new car price
• Achieved MAPE of 5%; Devised the method to give confidence on the predictions
• Exposed the model through flask API

Name Entity Extraction from Documents

• Built the Name Entity Recognition model to extract information such as Name, Father Name, Engine Number, Address etc from documents such as PAN Card, Aadhar, RC and Cheque
• Used RASA which is open source machine learning framework for chatbots to build the model
• Achieved precision and recall above 85% for almost all of the entities
• AWS Textract was used to convert image to text.

Recommendation Engine

• Recommending dealers to get them on the platform to auction on a used car
• Sequential features were developed by taking last 6 months of participation history of the dealer
• Deep Learning model was trained by creating positive and negative samples.

QC automation using Computer vision

• Trained Yolov3 model to mask vehicle number plate.
• Accepting and Rejecting images based on text or person present inside car boundary; Boundary box of person, vehicle and text are determined by Yolov5 and AWS rekognition API respectively
• Built image classification model to determine Color, Make and Model of a car with Inceptionv3

Personalisation Engine

• Built personalization engine to sort listed used cars based on user preferences.
• Engines comprises of two scores - One score is based on user browsing history (what car user has seen), other score related to car performance in terms of its views and leads
• Engines runs on more than 30k cars and its API gives response within 250 ms

Lending Model

• Built SMS based credit scoring model for a small ticket size loan distribution and predict probability of default;
• Used Extreme Gradient Boosting for training the model to predict probability of default;
• Carried out feature selection for the model to get final 42 features out of more than 200 predictors.
• Wrote complex SQL queries for feature extraction and data processing.
• Deployed the model by building API using Flask and processed 14,000 to 20,000 applications on daily basis
  

Wallet Model

• Developed credit scoring model using transaction data of a user collected from Mobikwik App; Applied Gradient boosting to predict the probability of default.
• Formulated UDFs in Hive and Pyspark to prepare data set for the training of the loan distribution model for 10 million users.

Supply Chain Optimization Tool

• Solved a constraint optimization problem of demand and supply using Pyomo and glpk solver by maximizing utilization of dispatching trucks.
  

Relay Model

• Business Problem: Addressed the existing inefficient way of material transportation on long hauls by converting two driver system in a truck to a unique relay model to significantly enhance miles covered per day.
• Algorithm: Formulated greedy algorithm to create the relay model for Long Haul Transportation in Python to optimize the schedule of drivers that involved minimizing idle time of the drivers
• Result: Optimal number of drivers and life cycle of every driver required to run the model.