Tushar Sanjay Khustule

Recommendation System for Songs (Music Player)

• Created user-song matrix mapping from unmapped songs across user due to unclean name using Fuzzy String matching
• Used Matrix Factorization to create embeddings of songs using Collaborative Filtering
• Created a hybrid recommendation system using above song embeddings from collaborative filtering and collected Audio Features of songs from Spotify's song metadata API
• Increased engagement time (song listening time) by 2.3%

AI Bot (Football World Game)

• Developed a bot for the two player based football free kicks game - where a player can play against a real player or bot
• Used MCTS (Monte Carlo Tree Search) to get an optimal action set
• Used another model XGBoost to choose exact values from particular action set generated by MCTS
• LSTM is used for Dynamic Difficulty Adjustment for the user based on previous actions and games of user which is used in MCTS to generate action set
• Deployed on AWS - using Flask, Gunicorn, Nginx and docker container for real time prediction
• Increased D1 Retention by 3.7%

Player Segmentation (Football World Game)

• Created session level features for the particular user using the gameplay logs, as well as engagement, purchase behavior
• Clustered mainly on the basis of skills using gameplay data to understand skill of the user
• Used Dimensionality Reduction techniques as PCA, t-SNE, UMAP and employed clustering models such as K-Means, DBSCAN, GMM on the top of reduced dimensions
• Achieved silhouette score - 0.63, DB Index - 0.83, CH Index - 9800+
• Resulted in better difficulty adjustment by changing bot setting and player matchmaking

Revenue Forecasting (Time Series)

• Enhanced MasterCard's AI-enabled Revenue forecasting engine. Used ensemble of various models including conventional time series models like ARIMA, Prophet, MAPA, etc along with Deep Neural based models like DEEPAR, Transformer, etc based on R and Python
• Enhanced this engine to incorporate post Covid disruption for ~600K products. Used scenario-based modelling by using models of differing capabilities like TRMF, DLM, LightGBM etc.
• Solved challenge of providing best forecast at any hierarchical level of products being solved by Hierarchy Reconciliation
• Achieved MAPE of 1.6% on global level beating forecasts by Mastercard's Enconomists and Financial analysts
• Managed stakeholders including regional CFOs of Mastercard
• Received 'Own It' award for Ownership of project
• Awarded with 'CFO Award' for the contribution and impact of the project in company

Time to Hire Prediction (for HR Dept)

• Efficiently Predicted time to hire a particular person
• Used XGBoost, CatBoost for the regression problem as data mostly contained Categorical values
• Achieved RMSE of 5.6 days, where average time to hire was 38 days

• Selected as Tier-1 consultant
• Created algorithmic trading models known as 'Alpha' for trading stocks in markets like Asia, Europe, North America, etc.
• Implemented Mean-Reversion, Behavioral Exploitative, US Market news based, as well as company specific daily stock price behavior and company balance sheet history based strategies.
• Created 120+ different alpha strategies for different markets

• Built plugin for Photoshop using Javascript and Objective C, that exports objects in Photoshop to various formats like 'JSon', 'XML', etc.
• Worked on basic UI designing and animation with 'Sketch' app on macOS