A python application to convert speech to sign language representation and vice versa for enabling those who are speech impaired to interact easily with others.

Build an android application that compares the best rates available for users from different e-commerce websites. It was also capable of price prediction by analysing data on previous price trends.

Developed an interactive android game to guess the jumbled words.

A web application to help companies manage inventory storage spaces.

a tool to automate the matching of JDs with resumes to get the top candidates out of thousands using FastText Skip-Gram Embeddings, Named Entity Recognition and tree-based search algorithm.

for US housing insurance domain problems related to Eave Height Calculation and reduced median RMSE from 30% to 16% resulting in a 70% reduction in potential penalty. Also, achieved a F1-score of ~98% and AUC ~95% for Storey Count.

for Measurement Automation and Accelerated Tie-Pointing.

leveraging deep learning to resolve queries for Airtels 400M+ monthly active users. Currently deployed at a commercial scale handling thousands of conversations everyday in English or Hinglish with an accuracy of ~85%.

constituting Face Recognition, Face Matching and Face Liveness Check models. Updated the platform with Face Segmentation, Face Blur Detection, Document Forgery Detection, Real-time blacklisting and Face Search to improve service efficiency and reduced manual effort by over 25%.

for FortressIQ equipped with an OCR solution that outperformed the accuracy of Google Vision OCR by 12% benchmarked on the ICDAR 2013 dataset. We built and trained custom models for region proposal and text recognition with semantic correction using BERT. Reduced overall cost from an estimated $1M to less than $50k per year.

from product title or description using word2vec, N-grams embeddings and kNN classifier.

using RNNs to parse guest comments on website, assess sentiments and classify into predefined themes, finally responding in a standard manner or directly to necessary systems to respond.

for sentiment analysis, Text-to-SQL and content generation.

Hand Gesture Recognition: Conceptualised and built a ConvLSTM model on TensorFlow. Used Bayesian Optimization techniques for hyperparameter tuning.Single-Shot Face Recognition: Used SSD Mobilenet for detection and Inception Resnet v2 for feature extraction. Introduced TP-GAN for photorealistic and identity, preserving frontal face view synthesis from any poses.Took complete ownership of this project and showcased it at CES 2019 in Las Vegas, USA with a live demo.