LLM model development

developing ML Ranking Models

Developed a novel solution for determining when models should be activated and deactivated to save CPU cycles through multivariable optimization.

Performed feature engineering and trained Random Forests, Gradient Boosted Trees, and Neural Networks to classify whether cellular devices were using 5G with 0.75 F1 score on highly imbalanced datasets.

Devised new ResNet-LSTM models to classify audio segments from multiple datasets with an average of 0.9 F1 score for FakeNetAI’s audio deepfake detector.

Deployed model on to a web server with AWS and streamlit backend.

Created custom unsupervised PyTorch word models with importance weighting and embeddings to achieve an F1 score of .75 trained on unlabeled data.

Published Self-Supervised Pretraining Improves Self-Supervised Pretraining to the WACV 2022 conference.