Harish Natarajan

Computer Vision Engineer

Aftershoot

04.2022 - Current

Contributed to the development of a sophisticated Retrieval Augmental Generation Pipeline and optimized Large Language Model (LLM) models using cutting-edge techniques like Parameter Efficient Finetuning (PeFT), Low Rank Adaptation (LoRA), and QLoRA. These improvements were tailored to efficiently handle bespoke inquiries within the Aftershoot Ecosystem, resulting in a notable 40% reduction in workload for the customer success team.
Engineered an internal image generation tool leveraging stable diffusion for model training purposes, subsequently deploying it on AWS Sagemaker for seamless integration. Additionally, conducted research and devised logic for monitoring Model Drift, facilitating timely model retraining in response to data drift occurrences.
Developed a deep learning Vision Transformer model for closed eyes and emotion detection, achieving 90% accuracy and capable of detecting individuals wearing sunglasses.
Researched and implemented machine learning solutions to address business challenges affecting 10,000 photographers.
Automated data preprocessing pipelines and model training workflows with Python scripting, reducing manual efforts by 50%.
Deployed diverse computer vision models including face detection, pose estimation, and eye gaze detection within the Aftershoot ecosystem.
Engaged in the development of a diverse range of deep learning models, including speech fluency detection, speech-to-text, and various others. Quantified these models into CoreML, OpenVINO, and ONNX formats, and subsequently deployed them within a Kubernetes cluster.
Streamlined data preprocessing and model training pipelines by automating routine tasks with Terraform Python scripting, reducing time spent by 50% on manual processes significantly.
Optimized cloud storage costs by 3x and established deterministic training pipelines for reproducibility.

Similar Profiles