Led a team to create a Retrieval-Augmented Generation based question-answering web application that references 3 million research papers using LangChain, Cloud Run, FastAPI, and PaLM 2. Also created a scalable pipeline to ingest documents and create embeddings using Dataflow, Vertex Vector Search, and BigQuery.For an online marketplace for buying and selling cars in the UK, led a team to create ML models to predict the maximum auction price for used cars using Vertex AI, Python and SQL. Improved over existing metrics by over 70%.For a globally recognized engine manufacturing company, spearheaded a project team employing object detection models like RetinaNet and YOLO to identify and predict defects in engine components using Vertex AI, Python and Sci-kit Learn. Enhanced the speed of defect detection by 2x.For one of the world’s leading dating apps, led a team to create ML models to predict the intrigue (or interest) two users have for each other using Tensorflow, SQL and Sci-kit Learn. Improved over existing methods by 30%.For a geospatial imagery company, led a three-person team to rewrite the training and inference pipelines for a computer vision model (originally written in Pytorch) to Tensorflow to reduce the training and inference time by 5x using Tensor Processing Units (TPUs)

Functioned as a consultant to a client that offers webcasting products to thousands of users. Aided the client in diagnosing errors and improving the running of their pipeline using best practices to generate live transcriptions using the Speech to Text API. Reduced by word error rate by 10%.Was a part of a team that created a scalable MLOps pipeline for preprocessing and transforming data; training, evaluating, and deploying computer vision models for a manufacturer of computer memory and data storage using Dataflow, Tensorflow Extended (TFX), AutoML, Cloud Composer, Kubeflow and Vertex AI.Was a part of a team that reduced the inference time by 10x of a high throughput DNA sequencing model for a California based company that uses novel electronic chips. This reduction was achieved by utilizing concepts such as Fire Modules and Depthwise Separable Convolutions that reduce the number of parameters of CNNs along with Nvidia TensorRT.

Led a team of engineers to build and deploy deep learning models trained on computed tomography (CT) scans to identify pathologies in the lungs for one of the largest pharmaceutical companies in the world. These models are currently undergoing FDA approval.Piloted a team of engineers to create and deploy deep learning models to detect the location of garments such as dresses, pants and handbags in a user uploaded image and filter out invalid images as part of a pipeline to construct a state of the recommender system for apparel for a fashion-tech startup.Helmed a team of engineers to build a deep learning model to identify in an image, various attributes of an apparel such as style, sleeve type, neckline type and length type for the above mentioned startup.

Created machine learning models to predict peer-to-peer (P2P) loan defaults based on data from two P2P lending platforms. More than $100,000 were invested by the client into notes based on such loans using only predictions from the models. Developed a model to predict student admissions and placements for a reputed IT Training provider. The model is currently used by the company to help it earmark students that are more likely to take admission and are easier to place. Built a model to predict which employees would leave a healthcare company and what are the factors that influence them quitting the organization.