Performed semantic segmentation to map farmland in satellite images. The process included cross-referencing open street maps to find images containing farmland using geo-cooridinates, utilizing QGIS software to create ground truth masks, and implementing several convolutional neural networks to train a predictive model. This was my final project at Metis.

Used natural language processing techniques to take an input of a block of text and output a set of recommended blog posts related to the subject. The process included programmatically building a corpus of several thousand blog posts and using a TF-IDF matrix to suggest posts using cosine similarity. The tool was built using Flask.

Used old school shallow learning techniques to classify insect images. The technique employed was "Bag of Visual Words" as inspired by the popular NLP technique. The process involved building a set of feature descriptor vectors for each image, building a codebook for all images using KMeans clustering, building a histogram for each image, and finally fitting a support vector classifier.

Built a simple regression model to predict baseball stats better than the baseline of using the previous season's stats. In fact, the only data that was used in building the model was data that was minimally affected by the natural randomness in the game. The idea was to prove a concept that the true nature of a player's ability is not inherent in a single season's statistics.