Jun 2022 – Present · 3 yrs 9 mos · Bangalore Urban, Karnataka, India

• Led Byaj Book from inception to a fintech platform with over 400,000 registered users,
• Led Comprehensive market research, identifying lucrative industry trends and uncovering innovative
• opportunities. Implemented Data Pipeline to segment and filter users and create funnels.
• Developed Flask API-based Aadhar and PAN card verification system. Implemented CNN for information
• extraction. Integrated OCR solution into Flask API for seamless document processing. Achieved more than 97% Accuracy on the Aadhar and PAN data extraction.
• Engineered a Liveliness Detection API to authenticate users. Utilized features from the VGG16 ImageNet model. Implemented Support Vector Classifier (SVC) for liveliness detection. Achieved 99.2% Accuracy on the Kaggle Dataset.
• Applied machine learning and deep learning techniques to develop intelligent features and predictive analytics, including identifying high-trust users, driving user engagement, and optimizing lending processes.

Apr 2021 – Jun 2022 · 1 yr 2 mos

• Non-Intrusive Speech Quality Prediction: As a part of the INTERSPEECH challenge worked on the speech quality prediction system, using a transformer-based encoder network before applying attention pooling. Secured 6th position worldwide and Research paper accepted in INTERSPEECH 2022.
• Speech-to-Speech Translation: As a part of the Automated Speech Recognition(ASR) team, worked on a speech translation problem involving the conversion of Korean audio signal features to English audio signal features. Using the Transformer model as the fundamental sequence-to-sequence architecture with the addition of auxiliary decoders to train on parallel tri-phone aligned data.

Jun 2019 – Mar 2021 · 1 yr 9 mos

• Far-Field improvement for Wakeup and ASR: Bixby is a voice assistant indigenous to Samsung smart devices. Built and deployed lightweight CNN-based architecture and improved the Far-Field wakeup detection accuracy in Bixby. Achieved exceptional results by attaining over 99% accuracy in wake-up word detection through innovative algorithm development and rigorous testing methodologies.
• Improved Multi-Head Attention For Single Channel Distant Speech Recognition: Distant speech recognition is a challenging task mainly due to background noise, reverberation, and distance-related degraded speech. We used stacking and sub-sampling of frames, changed the self-attention to improve the Query-Value interaction, and also introduced CTC-based loss. We got improved results on AMI meeting corpus than the underlying baseline

May 2018 – Aug 2018 · 3 mos · Bengaluru, Karnataka, India

• Conceptual Segmentation of the Questions for CAR: Segmenting questions conceptually into multiple query fragments could help question-answering (QA) systems in answering questions with complex information needs, Information relevant to answering a complex question is typically distributed across multiple content segments over multiple locations within the document collection. We Used a Knowledge Graph (KG) and deep neural model to decompose questions into fragments representing relevant sub-concepts and used those fragments to retrieve relevant content segments needed to infer the answer.