Ownership of AI-based replacement of Video compression In-Loop filter achieving 10% bd-rate gain
Curated data using quantization range resultant artifacts based binning for model generalization
Researched and developed a novel training strategy to train a smaller network better than a more complex network trained without our unique training strategy
To further reduce the model multiply-and-accumulation operations for feasibility of deployment in devices, we innovated on selection from amongst a plurality of models when deployed in video codec along with bitstream signaling optimization
Filed five patents based on the aforementioned approach
Part of the OCI Exascale team and was responsible for enhancing cloud-based Exascale services
Created and completely owned a FLASK-based web-service and deployment on cloud server
The Flask application is used for synchronous resource management on a server which enabled more efficient usage of the resource being managed by our service
Worked on Reverse Geocoding which upon receiving a set of coordinates predicts an address using Named Entity Recognition to create a custom clustering tree to generate a set of candidate addresses
Used beam search and reference data-based filtering we output a final predicted address. The project is currently in production in Amazon India marketplace
As the next project, created an Address Classifier involving responsible extraction of customer data by protecting privacy and ensuring to remove bias, towards more rich and developed areas, from the embeddings
Developed a multi-branch CNN architecture with the input branches mimicking n-grams. This model resulted in 6% better prediction AU-ROC than the previously used LSTM model
The Model is currently in production in Amazon middle- east marketplace
Researched deep learning methods to produce images with enhanced legibility upon zoom leveraging a multi-focal lens array system to generate images of different zoom levels
Created a custom-Unet model to fuse such images and produce a single image with better zoom legibility
Curated the entire dataset required and performed adequate data-preprocessing as the first leg of the pipeline
The solution achieved 1.5 dB Peak-Signal-to-Noise-Ratio (PSNR) improvements over the baseline method
