Nikhil Reddy

Developed advanced audio ML models to classify coughs and remove speech from audio data using techniques like non-negative matrix factorization (NMF), Hugging Face Transformers, and Wav2Vec, which improved performance and reduced complexity. Integrated new signal processing features (such as spectrograms and Librosa) and used knowledge from pulmonologists to enhance models for medical use. Led research on using cough analysis for COVID-19 screening, resulting in the publication of three peer-reviewed papers for high accuracy, speed, and reliability. Created models to identify respiratory diseases by analyzing acoustic cough signals, leading to impactful research publications. Designed scalable model training and deployment pipelines with Amazon SageMaker, utilizing S3 for efficient data storage and Git for version control ensuring high performance and scalability.

Developed and deployed audio-based cough classification models (risk, pattern, disease, TB) using TensorFlow/Keras, showcasing skills in audio feature extraction and machine learning. Created a novel cough vs. non-cough classifier, enhancing the company’s core product offerings. Collaborated with pulmonologists to integrate medical expertise into models and published medical validation for product launch.

Worked on revolutionizing aircraft engine monitoring for DRDO. Developed a machine learning framework for the Kaveri engine to reduce the number of sensors needed while maintaining prediction accuracy. Implemented anomaly detection algorithms and trained classification models. Conducted analysis on sensor data to detect damage and identify damaged components. Reduced features from 1000 to 2 in the aircraft test-run dataset using alternating iterations of heat-map correlation elimination and random forest regression, achieving 96% prediction accuracy. Performed prediction by reducing parameters from 6 to 3 using linear regression and random forest regression, achieving 90% and 98% accuracy, respectively. Worked in Systems Engineering and performed performance evaluations for proposed solutions. Performed classification on sensor fault data using deep learning techniques such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), achieving high accuracy.