Weekend Hacks

Static Portfolio made with Ionic (Angular) and hosted on GitHub Pages

• Javascript
• Angular
• Ionic
• HTML
• CSS
• Bootstrap

Finding a subset of Troops to upgrade, which together work across most levels of the game

• Python
• Dynamic Programming
• Data Cleaning

Figured out how to convert Jupyter Notebooks into Python scripts using Jupytext and putting them under version control. Because Jupyter Notebooks are JSON files containing outputs of the last session as well, and diffing them doesn't make much sense. This along with git tags helps in tracking the evolution of experiments.

• Git Hooks
• Jupytext

Performed Preprocessing to fill null values and trained a Random Forest Classification+Regression Ensemble Model to predict Price. Final OOB R2 score was 0.97. Insight: Data-visualization showed that the number of models in market decreases 10 fold as the price increases 10 fold. Coincidentally, adding Price Bracket as a feature improved R2 score.

• Python
• Scikit-Learn
• Pandas
• Feature Selection by Feature Importance
• Exploratory Data Analysis

Finetuned a VGG network to classify Cloth Patterns. The image features were extracted from all the max-pooling layers of VGG19 and then concatenated and fed to 2 dense layers.

• Python
• Pytorch
• Data Version Control
• Google Collab

Named Entities are usually not in the dictionary and can be comprised of multiple words.

• HMM
• CRF
• SVM

The primary step in NLP is converting text into vectors. For this we use Bag of Words. To keep the size of BoW manageble we need to reduce the size of dictionary. To do that we replace words with their stem.

• Python
• Coursework

Participated in a Kaggle challenge to classify Poems into different categories. Combined word vectors to create sentence embeddings and trained different statistical models to predict the categories.

• Word Vectors
• Sentence Vectors

Practising Mental Math is boring if you have to erase do the same problems again. I created a simple random problem generator and GUI which automatically checks the answers and shows the time taken.

• Python
• Tkinter

As the dataset was quite large, obtained image features using a pre-trained CNN and trained a smaller network on top of the extracted features.

• Python
• Bcolz Array
• Multi-Threading