Projects
A Kaggle challenge to identify, from a photograph of cassava leaves, which of four diseases is afflicting the plant or whether it is healthy. Training data was over 21,000 images crowd-sourced from farmers — heavily imbalanced and extremely noisy. Using PyTorch, I performed image augmentation and used a 50-layer residual network to reach over 70% accuracy on the test set, an 8.5% improvement over predicting the most frequent label.
A team effort on a UNICEF challenge to predict Human Development Index from country-specific data. I contributed ETL in Python and pandas over several large CSV files, aggregating related data and generating new data points by mapping independent but correlated sources — a 30x speed-up over naive iteration. I also contributed the loading and scaling of that data in D3 and Django.
A group project in MATLAB implementing an agent to play Go. For the game itself I contributed the recursive logic that efficiently determines when pieces are captured. For the agent I built a neural network from scratch, along with a genetic algorithm.
Connected an Arduino Uno, a motorshield, and a sonar to a Snap Circuits rover, then programmed it to drive autonomously while avoiding objects.