Ravi Kolluri

Accomplishments: Mentoring a team of 4 Data scientists to build data science practice from scratch for the group. Speaking with various stakeholders to bring product ideas to fruition.

Developed a deep learning model for classifying incoming news articles into respective categories (Business, Entertainment etc..). Implemented an XLnet model with a pre-trained BERT layer and fine-tuned to achieve an overall accuracy of 87%. Used Named entities (Person, Organization and Location) from News articles and User profiles to do targeted recommendations to improve App usage and Customer Experience. Developed an Encoder-Decoder deep network with attention mechanism model to translate titles in Walmart's catalog in English to Spanish and French. Utilized the same architecture for candidate generation using title generation on Competitor's catalog. Impact of the project: 13% improvement in match coverage of products on Amazon. All the catalog images on Walmart.com are stored as embeddings, Locality sensitive hashing is used to do reverse matching to help in an increment of 4% in product match coverage. All candidates during matching are sent through an ML Model for computing a final probability of Exact match of product. We compute sentence similarity (Using a transformer architecture trained on existing matches), image similarity (Open VGG16 network and Cosine Similarity) and other entities are extracted to send as an input to the ML model. We achieved a precision of 95% with a threshold of 0.90. Responsible for mentoring three Data Scientists, Contributed to an overall improvement of match coverage by 18% for walmart.ca.

Pre-match player projection: Developed a model to predict the pre-match performance of each player using historic scorecards with a MAE (Mean Absolute Error) of 6 runs. An XG-Boost regression model was used on features derived under performance, form and streak buckets. Live Team Projections: Using batsman's average score as mean of the distribution and strike rate to determine the resources, we run a Monte-Carlo simulation of the match for 10000 times per ball. We derive win probabilities and also expected scores at each moment of the game for any format assuming a normal distribution of the sample scores to calculate win percentage of respective teams. Automated Fantasy XI: We set up an integer programming model to maximise the utility and arrive at a best possible combination of a team for fantasy 11. Accomplishments: Mentored a team of Two data scientists to deliver Final four teams prediction model for a round robin format of matches and other use cases on user analytics and engagement. Fintech For the Poor: Co-authored a research paper in collaboration with Omidyar networks to study the impact of mandating AADHAAR seeding to bank accounts and AADHAAR ENABLED PAYMENT SYSTEMS(AEPS) towards a goal of achieving financial inclusion in India primarily studying the Business correspondent model in India. We used empirical models to show introduction of AEPS has improved the inclusion of more people in the formal financial sector. Mutual Fund Selection: Developed a Model which uses an ensemble of a K Means Clustering, Additive seasonal ARIMA model to forecast the return for various time periods. Various features like risk, quantitative attributes of the fund manager, Sharpe ratio, and Sortino ratio are used to select the top 3 mutual funds in each asset class. A Monte Carlo simulation of model portfolios was generated to identify the combination of funds that will maximize "alpha" and minimize "Beta" of the entire portfolio. Portfolio Optimization: Optimization of weig