Experience
Machine Learning Engineer (Part-Time)
Created a LightGBM based Jewellery purchase prediction model to promote EMI based Jewellery purchase scheme for Kalyan Jewellers' customers. LightGBM was initially trained on user features such as income, city, previous purchases etc. over a year to achieve an accuracy of 95%. In deployment, the model also updates this score based on every user activity such as a new purchase from any branch. On performing a t-test, the model is expected to increase jewellery purchase for the year 2023 by 15%.
Machine Learning Engineer (Part-Time)
Developed a website tool to generate creative bios with fine-tuned GPT2 based on the users' profile in English and mother tongue in collaboration with the technical team at KalyanMatrimony.com. The feature was integrated in the existing website by deploying the GPT-2 with Fast API using EC2 instance, launched and managed on Amazon Elastic Container Service with Auto Scaling. This feature was provided as a campaign to users to generate a summarized creative bio for their profiles. The feature resulted in a 20% increase in profile views for users evidently increasing customer satisfaction.
Machine Learning Engineer (Part-Time)
Created an ensemble model using Graph boosting and LightGBM to increase accuracy in subscription prediction using user features such as login_count, message_count, income etc. from 96.5% to 99%. The Subscription prediction-score is calculated on a parallel voting basis which increases AUC from 0.95 (individual models) to AUC 1.0 (ensemble) when trained on 20000 users' activity over six months. The projected increase in sales is about 20% for year 2022-2023
Machine Learning Engineer (Part-Time)
Developed a subscription-prediction model for users in KalyanMatrimony.com. Deployed Graph boosting model (XGBoost) which uses user features such as login_count, message_count, income etc. over a year to achieve an accuracy of 96.5%. In deployment, the model also updates the subscription-prediction score based on every user activity on a per-day basis. For the year 2021-2022, this model has increased tele-sales in the company by 40%.
AI/ ML Intern
Developed and created a real time ID verification process from videos using video enhancement and the U-Net Deep Learning model with 89.8% accuracy. Rapidly prototyped new data processing capabilities in video enhancement using Python to confirm integration feasibility into existing systems. Collaborated with multi-disciplinary product development teams (AGILE-Jira) to identify performance improvement opportunities and integrate trained models.
Deep Learning Intern
Identified new problem areas in the check verification system and identified check frauds and errors using transfer learning on a custom synthetic dataset (VGG16, VGG19, 96% acc). Also incorporated signature verification using thin-plate spline warping for spatial transformation into existing product "SnapCheck".