Akila Thalgahagoda

Implementation of a deep learning framework to predict the correctness of the action of a swimmer from a video. Finetuned multi-modal VLMs using Parameter Efficient Fine-tuning techniques (LoRA / Adapter Tuning) to generate captions on the swimmer pose. Working on the end-to-end implementation from problem formulation, research, development, data set management, and deployment.

Video Surveillance R&D: End-to-end ownership in ML features of a video surveillance product that used a combination of deep learning and traditional computer vision. Implemented custom model architectures and training recipes for relevant deep learning architectures from academic publications. Trained, improved, and debugged model (eg., CNNs, VisionTransformers) training pipelines. Data set management: data collection, using available deep learning models (eg., object detection, segmentation, and generative models) and foundational models to clean/enrich the existing data. Model compression (eg. Knowledge Distillation) for deploying in resource-constrained environments. Model conversion to ONNX/Caffe and integrated to the product SDK. Implemented MLOps best practices for experimentation management using MLFLow. Maintained an experimentation framework based on Hydra for config-driven experimentation. Some projects worked: Human Action Recognition (pose classification), Clothing Identification, Object Color Naming.

Customer Churn Prediction/Segmentation: Worked on the end to end data science project life-cycle; extracting data in disparate databases, deriving features, model building, testing, deploying, and presenting the results to business stakeholders.