May 2022 – Jan 2024 · 1 yr 8 mos · Bengaluru, Karnataka, India

• Instrumental in the training and implementation of Multi-gate Mixture-of-Experts(MMoE) models to augment user engagement metrics, successfully boosting Day-1 retention by an impressive 0.7%.
• Conceived and executed innovative surface-specific models, leading to robust experiments at Sharechat. This pivotal initiative culminated in developing and integrating a unified Catboost model within the Suggested Feed ranking system, resulting in a notable 0.5% uplift in user retention.
• Enhanced feed quality by developing and refining various heuristics for the Final ReRanking layer utilising alpha-ndcg. This strategic optimization drove a measurable 0.2% increase in user retention.

Aug 2020 – Apr 2022 · 1 yr 8 mos · Hyderabad, Telangana, India

• Developed multiple deep learning models spanning across Computer Vision, Natural Language, and Speech
• Developed a new model architecture for information extraction from documents using Yolov4, OpenPose, and LayoutLM; responsible for the whole process including data collection, annotation, model training; resulting in increased revenue of 5M (5000 + Documents)
• Developed a 3D CNN model to identify and localize violence on live videos using C3D and Self Guided Attention along with GradCam for localization; achieved 94% accuracy on the test set (100k + clips)
• Led a team of 4 people to develop a Transformer based model to identify crime-related tweets and classify them using DistillBert and zero-short learning to label the tweets; achieved 92% accuracy on the test set (5000 tweets)
• Developed a speech sentiment analysis model that gives real-time feedback to the customer service personnel using an LSTM architecture; achieved 90% accuracy on the test set
• Developed a Pose Monitoring and Pose correction system for baseball pitchers using PoseNet, OpenCV, and Dynamic Time Warping

Jul 2019 – Jun 2020 · 11 mos

• Automatic Speech Recognition:
• Lead a team of five people, working on the prestigious MHRD project to generate NPTEL video subtitles automatically with only 12% WER. (Word Error Rate)
• Developed multiple methods of using language models like GPT-2 in ASR that helped train good quality speech recognition models using less data and reduced the WER by 4%

May 2019 – Jul 2019 · 2 mos · Bengaluru Area, India

• Led a team of 2 people to develop a scalable real-time prediction system that brought down severity level-three errors by 50% using the Seasonal-ARIMA model.
• Built the entire system starting from fetching Data from ELK, Data processing, Model development, and dashboard for visualization, serving 40,000 servers

Dec 2018 – Jan 2019 · 1 mo · Greater Chennai Area

• Research Internship at UGV (Unmanned Ground Vehicle) laboratory, CEG,Anna University.
• Developed a hybrid system that uses both LIDAR sensors and CNN to detect and avoid obstacles, worked on creating our dataset, and trained custom CNN architecture to achieve 98% accuracy over 8000 images

May 2018 – Jul 2018 · 2 mos · Greater Bengaluru Area

• Lead a team of two people to implement various Deep learning models using ESPnet to improve the team’s speech-to-text capabilities.
• Our best model CTC (Connectionist Temporal Classification) + Multihead-Attention + Label-Smoothing hybrid has achieved a WER (Word Error Rate) of 11.5% and CER (Character Error Rate) of 4.7% on the test (clean), Librispeech (960hr) dataset.

May 2017 – Jul 2017 · 2 mos · Greater Chennai Area

• Worked on Image Generation using GANs(Generative Adversarial Networks)We used GANs to generate images which were useful for the organization. We generated completely new images and also improved the resolution of the existing images using GANs.
• The main take away of this internship was that I learned a lot about Deep Generative Adversarial Networks and a lot about the practical machine learning.