Python

Built a CNN with transfer learning to classify nutrition data. Applied image recognition to health analytics and pitched an app concept with a business plan. – The “AI nutrition grader” builds upon “Presenting Nutritrack,” a project that won my team an AI app design pitch competition when we presented a custom CNN proof-of-concept and discussed our business plan. The model built for this project, unlike the Nutritrack custom CNN, is a transfer learning design. The source for the data is the Nutrition5K dataset (RGB images of food on white plates). This model includes an augmentation layer to increase the size of the dataset, by changing the angles of the plates to create a greater volume of training data. For transfer learning, the AI nutrition grader CNN model uses a MobileNetV2 backbone pretrained on ImageNet. The model builds a custom classification head for feature extraction and then fine-tunes the deeper layers to adapt to its own food image dataset (drawn from Nutrition5K). The README for this project on GitHub provides access to the images selected for this project’s dataset, with details about how to download large image files so others can clone this repository and replicate the AI nutrition grader project without needing to download the full Nutrition5K dataset.

Sep 20, 2025

Analyzed 780+ graded mathematical proofs to compare LLM vs. human grader consistency and accuracy. Identified feedback errors and applied statistical testing to evaluate grading reliability.

Sep 10, 2025