Experience

Fi, Data Science Intern

Implemented Hierarchical Classification as a solution to the problem of Transaction Categorisation by using pre-trained BERT embeddings as feature vectors for the transaction remarks, and a Logit pipeline to classify the transactions into different bins hierarchically.
Preprocessed, cleaned and labelled the production data by writing automating scripts.

Worked on the problem of applying Reinforcement Learning on Neural Machine Translation (NMT)
Built upon the work of Tsz Kin Lam et al. on Reinforcement Learning Approach to Interactive Prediction Translation
Majorly focused on using latin and non-latin language pair datasets such as Eng-Viet for our experiments.
Experimented with different reward functions such as BLEU, YiSi, CharF by analyzing reward scores and translation quality to figure out which would work the best under given conditions for the problem of RL on Interactive Prediction Translation.

Worked on the GST (Indirect Tax System of India) software which helps SMEs, Chartered Accountants and Enterprises to file GST returns and simplifies their filing experience
Built new features for the GST product using frameworks and libraries like NextJS, GraphQL, React, Apollo (Client), Jest

Optimized neural network architectures for Single Image Super Resolution (SISR) to upscale images from low resolution to high resolution by a factor of 2x or 4x.
Implemented the paper Wide Activation for Efficient and Accurate Image Super-Resolution (WDSR)
Achieved SoTA results on Div2K data. Clocked a PSNR of 30 on validation set using Python and TensorFlow 2.0