Developing state-of-the-art language models for Cash App's customer support team. Working on use cases ranging from ranking and classification, to text summarization, RAG and more.Tech Stack: Python, Kotlin, Hugging Face, Docker, Google Cloud, AWS, Prefect, Datadog.

At the intersection of software engineer and data scientist, focused on end-to-end NLP (Transformers, LLMs) and Computer Vision problems in the Infrastructure space - from requirements gathering to model training, deployment, and monitoring.• Trained and deployed an XLM-R text classification model to automatically handle non-actionable emails from customers, saving ~4,150 hours of network technician time/year. Used active learning to overcome the lack of labeled data.• Developed a custom BERT-based Named Entity Recognition model to extract maintenance event details such as location, time, and company, removing the need for complex regex and manual parsing.• Developed a pipeline involving a combination of object detection (Mask R-CNN) and OCR models to automatically detect barcodes on server racks, removing the need for manual asset tracking.

Intern from Jan 2020 to Jun 2020, converted to full-time Jun onwards. Working with product and full-stack teams to make customer-facing machine learning and deep learning models. Tech Stack: Python (PyTorch, pandas, NumPy, Flask), Docker, MongoDB, Google Cloud, AWS.• Named inventor on 4 patents for unique applications of deep learning techniques to the field of dentistry. US11189028B1, US20200411167A1, US20210358604A1, US20210357688A1.• Wrote a CNN + LSTM model for written handwritten character recognition, allowing parsing of thousands of documents that open source packages were unable to accurately parse.• Developed a duplicate image detection pipeline by training a Siamese network with triplet loss.• Deployed a resnet50 image classification model to determine image modality (xray, claim form etc.), achieving 98% weighted F1-score despite significant class imbalance.• Improved accuracy of claim form matching by 2% by using TFIDF on tri-grams of document id strings and using k-nearest neighbors to find the closest match.• Decreased run time of image processing code from 7 days to 5 hours. Key improvements were making MongoDB queries more efficient and parallelizing Python code using multiprocessing.

Responsible for building predictive models and leading complex analysis to improve Supply Chain efficiency. Tech stack: Python (pandas, NumPy, scikit-learn), SQL.• Saved $300k in transportation costs by creating a linear programming model in Python to determine the most cost-effective mode to deliver products to stores. Responsible for the end-to-end process including creating the data cleaning scripts, formulating the model and creating reporting for end-users.• Developed a random forest model to predict estimated time to unload a trailer.• Used k-means clustering to determine vendor region for new vendors.• Predicted expected sales for new products by creating a linear regression Pipeline in scikit-learn.• Decreased product spoilage rate by mapping transportation routes against external temperatures in Tableau. Gathered and cleaned the data using SQL and pandas.

Responsible for identifying process improvement opportunities in shipping, receiving and processing operations of distribution centers.• Performed statistical tests like t-tests, ANOVA, to validate process improvement ideas