Project Associate at the Indian Institute of Science, Bengaluru, India, with experience in federated learning, machine learning pipelines, predictive maintenance, and on-device inference. Also has experience as a Teachin…
Project Associate at the Indian Institute of Science, Bengaluru, India, with experience in federated learning, machine learning pipelines, predictive maintenance, and on-device inference. Also has experience as a Teaching Assistant, Research Intern, Software Engineer Intern, and Co-founder, Machine Learning Lead, with publications and projects in speech-to-text, TinyML, IoT, and embedded edge systems.
Developed an end-to-end Federated Learning framework for edge devices in speech-to-text tasks.
Devised a novel client selection algorithm employing the bandit setting of Reinforcement Learning to bring forth fairness in client selection.
Developed two new evaluation metrics because Word Error Rate used in traditional evaluations of speech-to-text models is inaccurate.
Created Heval, a weighted combination of semantic distance score and non-keyword error rate.
Created SeMaScore, which employs a segment-wise mapping and scoring algorithm and leverages both the error rate and a more robust similarity score.
Created a novel methodology for training large models on mobile devices by optimizing resource utilization based on mobile phone capabilities such as RAM and battery percentage.
1. Federated Learning
• Developed an end-to-end Federated Learning framework for edge devices in speech-to-text
tasks.
• Devised a novel client selection algorithm employing the bandit setting of Reinforcement Learning to bring forth fairness in client selection.
2. Evaluation metrics
• Since the Word Error Rate used in traditional evaluations of speech-to-text models is inaccurate, we developed two new evaluation metrics.
• The first metric Heval is a weighted combination of semantic distance score and non-keyword
error rate.
• The second metric SeMaScore employs a segment-wise mapping and scoring algorithm and
leverages both the error rate and a more robust similarity score.
3. On-Device training
• Created a novel methodology for training large models on mobile devices by optimizing resource
utilization based on mobile phone capabilities, such as RAM and battery percentage.
Indian Institute of Science, Bengaluru, India
4. TinyML
• Created an end-to-end machine learning pipeline for Predictive Maintenance, including fault
detection and remaining useful life estimation of a solenoid valve.
• Successfully deployed trained models on edge device (Arduino Nano 33 BLE Sense) for real-time on-device inference.
Assisted in conducting the lab sessions for the course “Design for Internet of Things” taught by Dr. T. V. Prabhakar.
Assisted in aspects of IoT protocols and micro-controllers.
Guided students in the completion of mini-projects.
Created an end-to-end machine learning pipeline for Predictive Maintenance, including fault detection and remaining useful life estimation of a solenoid valve.
Successfully deployed trained models on edge device (Arduino Nano 33 BLE Sense) for real-time on-device inference.
Developed a Bi-LSTM model based on textual input for a multi-class emotion classification task.
Expanded the project’s scope to accommodate larger contexts such as groups of sentences.
Provided performance analytics to pilots using flight operational data as a part of pilot assessment training.
Analyzed 1646 files of real-life flight data with each file consisting of 4196 parameters.
Developed algorithms for the identification of two new Non-Flight Operations Quality Assurance events.
Assisted in conducting the lab sessions for the course “Design for Internet of Things” taught by Dr. T. V. Prabhakar.
Assisted in aspects of IoT protocols and micro-controllers.
Guided students in the completion of mini-projects.
Currently directing the machine learning team for predictive maintenance of the fish tank.
Currently leading the machine learning team on a proof of concept for fruit quality assessment.
نظام التوظيف التشغيلي المعتمد على الذكاء الاصطناعي أولاً، الذي يوحّد ويؤتمت دورة التوظيف بالكامل للمؤسسات الحديثة حول العالم.
د.جوب منطقة حرة ذ.م.م. 2026 © جميع الحقوق محفوظة
