I handled the entire process end-to-end: dataset preparation, normalization, and augmentation; exploratory data analysis and feature correlation research; model selection, training, and evaluation; and deployment of a server-side inference system on client hardware.

The final system enables consistent classification of shrimp by size, making downstream sorting and processing significantly more reliable.

Shrimp Weight Estimation is a computer vision system designed to classify shrimp into size categories based on images.

The goal of the project was to automate the sorting process by replacing manual estimation with a consistent, data-driven approach.

Due to strict NDA constraints, only high-level details can be shared.

The work focused on the full ML pipeline: from dataset preparation to deployment. I worked with raw image data, performed normalization and augmentation, and analyzed feature distributions to better understand correlations between visual patterns and target classes.

Multiple machine learning approaches were explored and compared before selecting and training models that provided stable and consistent classification results.

A key constraint was infrastructure. Instead of relying on expensive cloud GPUs, the system was designed to run inference on the client's local machine equipped with GPUs, while still being accessible remotely.

This resulted in a hybrid setup where processing happens locally, but can be triggered and used from anywhere — balancing cost-efficiency with usability.