Data Scientist and AI Developer in New York, NY, United States

A complete question answering application for extracting answers from textbooks. Modern information retrieval techniques are successful in retrieving information from smaller documents. However, when it comes to larger documents, current options fall short.

This repository attempts to solve the problem of performing question answering on large documents. This requires a two-part approach. In one part, ALBERT is trained on the Standford Question Answering Dataset (SQuAD) QA dataset. In the other, we fragment a textbook into multiple sections using a rule-based approach. We can then compare user question embeddings to the embeddings of the sections to find the most relevant section(s).

I am the sole contributor—from product conceptualization to deployment—and the repository is currently in an MVP state.

Implementing the current state-of-the-art NLP model in fast.ai. The concept of transfer learning is still somewhat new to NLP and one that is growing at a very rapid pace. A model such as RoBERTa performs incredibly well on the SuperGLUE benchmark across several varying NLP tasks. This project facilitates the usage of RoBERTa with fast.ai.

I am the sole developer—from conceptualization to complete cross-integration—and the integrated model is available for use.

Thirty percent of users fill out psychometric surveys falsely. This is a problem for organizations that expect valid survey results for their survey expense. This project creates multiple models to find an approach that outperforms current validation methods. In the end, the aim is to reduce survey costs by at least 30%.

This project was built by a team of eight. I took ownership of building the complete pipeline for our LSTM approach, which yielded 80% accuracy and an F1 score of .76 on the validation set. The end deliverables are model weights that can be used locally to test predictions. Future goals for this project are to create an API for the LSTM model, which can be sent requests to identify false survey responses.

A deep learning computer vision model to detect fights in videos. By using five to ten frames from a two-second sample of frames, we extract features using a residual neural network (ResNet) model and then passing the extracted features to an LSTM (trained from scratch) to classify whether a fight is occurring in a video. My model is able to predict whether a fight is occurring with 90%+ accuracy on a balanced dataset. The accuracy is 71% on surveillance camera footage.

I am the sole contributor. The core development phase is complete and the next step is deployment.

A text generator web app built in under 50 lines of Python, using PyTorch. Within PyTorch, we use the transformers library to import the pre-trained OpenGPT-2 model. Secondly, the PyViz Panel library is used to fully create a web app, using just Python.

I am the sole contributor to this app. It is complete and intended to educate others on building complete text generation applications.