Enrique Balp-Straffon

An automatic screening/pre-diagnostic system for diabetic retinopathy using an ensemble of deep neural networks followed by a random forest classifier.
The model was trained to explore a vast space of convolutional neural network architectures inspired by Inception and ResNet (residual neural network) using best practices such as transfer learning, data augmentation, regularization, dropout, ensembles, and parallel training in several GPUs.
The system was designed to screen patients and take the workload off from ophthalmologists. The model had a sensitivity of 95% (true positive rate), while it had a specificity of 65% (true negative rate). This meant that the ophthalmologists only had to manually review 35% of the negative cases, resulting in much more efficient use of their time.
The system has not yet reached the stage of commercial distribution due to funding and regulatory issues.

I lead a team of data scientists and engineers in the creation of a tool to model oil price differences between locations in the USA. Based on historical data of production by thousands of wells, demand from refineries, and transportation costs in pipelines and trains, we used network analysis (networkx), geographic information systems (geopandas) and linear optimization to understand and visualize the price equilibrium of the system. The project was for a midstream company in Colorado, and was used to inform traders decisions.