Built a system for retrieval of food recipes given images of corresponding food

Beat the CCA baseline top-10 recall for recipe retrieval in the original im2recipe paper by 20 percentage points by using ResNet and BERT feature extractors and introducing cross-modality through a shared embedding layer

Implemented a second approach using triplet loss trained neural networks and attained median retrieval rank of 1 and top-10 recall of 82.49% for 1,000 random food images

Developed algorithms for building YOLO model bagging and boosting ensembles

Obtained an average precision of 87.5% on the Flickr-32 dataset using a generic logo detector system of two boosted YOLO models

Built a movie recommendation system leveraging user conversations, critics data and domain adaptation techniques, which is a re-implementation of this paper

Tuned hyperparameters for three CF approaches: KNN, SVD and SVDpp to obtain a 3% improvement in results

Experimented with neural MF and obtained comparable results of RMSE=1.232 and MAE=0.9569

Reproduced results for open set logo detection from the paper here achieving a 24 percentage point increase in mean average precision (mAP) compared to the original using YOLOv5

Focused on classifying textual logos and obtained a classification accuracy of 22.56% against 47 classes of the Flickr-47 dataset using a logo classification architecture consisting of YOLOv5 and template matching

Built a 11-layer deep autoencoder neural network using residual connections that colorizes black and white images

Trained the network on 10,000 images from FloydHub and deployed online via Streamlit

Analyzed a 55-million-record strong taxi fare dataset to determine varying trends in taxi fares across both location and time

Performed feature engineering and zoomed in on trips to and from airports and across different boroughs of NYC

Predicted taxi fares to an RMSE score of 4.28 by training a Random Forest model on the augmented dataset

Developed and deployed a customizable application that uses pandas and Exploratory Data Analysis to suggest an optimal team to be entered into the Fantasy Premier League fantasy soccer game

50+ monthly active users. Ranked top 2% in worldwide ranking among 8.2 million players in the year 2020

Built a text simplification system that can work on text and simplify it by removing difficult-to-understand words

Modeled and trained Transformer models that internalized the semantics of and recognized complex words in input

Improved the performance of the application by preceding the transformer architecture with a Complex Word Identification (90.23% accuracy) model that flagged the complex words beforehand

Determined the ages of abalones (snails) using classification techniques and leveraging their physical characteristics

Improved the accuracy of determining age using regression techniques and obtained a MAE of 0.936

Concluded that the dataset is not large enough to get the desired MAE of 0.5 implying correct age prediction

Expanded the ‘Space Invader’ game to include three modes of play: Arcade, Timed and Survival

Developed an Android application that acts as a combination of a reminder app and a notes app

Published the app on Google Play Store, and currently has 50+ installs with a rating of 4.6