Adding transformer-based NLP techniques and a graph neural network to tabular ML to achieve state-of-the-art performance (>90% accuracy) on a fine-grained credit rating prediction task

Paper | Code

Implements, visualizes, and compares three different time series forecasting methods to predict revenue losses from McDonald's ice cream machine outages. Pipeline scrapes data, retrains models daily via AWS

Forecasts | Code

Used named entity recognition and entity resolution to extract mentions of other companies from quarterly earnings calls, then constructed network data for interactive visualization

Full Network | 50% Sample | Code

A retrieval augmented generation chatbot that draws upon the TOSDR (Terms of Service Didn't Read) corpus to answer your questions about online privacy

App | Code

Quantifying the cost-effectiveness of prompt engineering via experiments. Conducts LLM conversations programmatically, finetunes models to evaluate text passages, and introduces cosine similarity measures and other metrics to compare model performance

Paper | Code

Generated contextual sentence embeddings (1,024-dimensional, several GB of data), loaded them into the vector database Milvus, and benchmarked retrieval efficiency

Paper | Code

Using classical and neural image embeddings with standard ML algorithms and finetuned end-to-end networks to achieve >93% accuracy on a vehicle type classification task

App | Paper | Code