I am a research affiliate at ETH Zürich working under the supervision of Dr. Leland K Werden on building AI systems for tackling
Climate Change. My interests lie within the broad areas of trustworthy ML and human-AI interaction on building safe, robust, interpretable, and interactive ML models for high-stakes domains such as healthcare, business and climate change.
Research Affiliate
| Assisted Forest
Regeneration
Dec '22 - Jan '24
Working under the supervision of Dr. Leland K Werden at the Earthshot lab on vision based models for estimating ground biomass, carbon sequestration, and reforestation cover of assisted forest restoration projects using remote sensing and GIS data.
Associate Researcher | Tech Mahindra
Aug '22 - Oct '23
Worked under Dr. Jibitesh Mishra and Ipsit Misra at Tech Mahindra's Center of Excellence of Artificial Intelligence. We developed various deep-learning models for smart traffic systems and intelligent healthcare systems. Additionally, I also taught an introductory undergraduate course “401-Deep Learning” to over 50 students across various backgrounds.
Research Intern
| Xu Lab
Aug '22 - Sept '23
Worked under the supervision of Dr. Min Xu at his lab. We developed an unsupervised clustering model for structural pattern mining of cryo-Electron Tomography (cryo-ET) data to sort and visualize marco-molecules without labels. Also contributed to a project focused on macromolecular structure classification using cryo-ET images with self-supervised learning. [
Code ]
Research Volunteer | Neuromatch Academy
July '22 - Aug '23
Junior Machine Learning Engineer
| Omdena India Chapter
July '22 - Aug '22
Worked as a task lead for the modelling and deployment teams to develop vision models for sorting and segregation of waste. We tested various state-of-art neural network architectures and hosted them on Hugging Face platform. Also presented a proof of concept to the UN SDG India team.
ML Intern | iNeuron.ai
Jan '22 - Apr '22
Worked at the intersection of data science and business intelligence on deep learning models for cardiac disease analysis and presented a data analysis report for business modelling team. [ Code ][ Slides ]
Summer Research Intern
| Centre for Visual Information Technology
July '21 - Jan '22
Developed a bilingual app featuring Ensemble Methods for First Action Prediction System (FAPS) to estimate cardiac disease risk using non-medical inputs like height, weight, age, health history, sleep, alcohol intake etc. with a real-day accuracy of 91.24%. Additionally, this application also provides details of nearby hospitals, clinics and doctors for public usage. This work was awarded the First Prize worth 2500 USD among 1000 teams nationwide at the 2022 Smart Odisha Hackathon.
Vision-Based Models for Sorting and Segregation of Waste
This work was completed during my tenure at Omdena. We developed object classification and detection-based models to segregate trash into common categories. Furthermore, we introduced a novel dataset comprising more than 16,000 images of trash items upon which evaluated multiple CNN and SSD architectures to achieve an overall accuracy of 96.14% for 10-class classification.
Satellite Data-based Pollution Forecasting using CNNs
In this research, we studied the feasibility and dependability of employing CNNs for predicting the BreezoMeter Air Quality Index (BAQI) using satellite imagery. We also introduced a novel BAQI dataset with over 10,000 satellite images and corresponding BreezoMeter air quality data from 56 cities and achieved an overall accuracy of 87.14%. This study indicates that cost-effective and widely available satellite images are efficacious for accurate air pollution forecasting.
CoughVid: Covid-19 Detection from Cough Voice Samples
As part of the Pfizer Digital Medicine Challenge, we devised an RNN and MFCC feature-based model to identify Covid-19 from cough audio samples. In real-world tests with over 320 patients, our model achieved an accuracy of 82.42%. We find that audio signal spectrograms can serve as image datasets for constructing models to classify audio samples.
We developed a transformer architecture for 3D medical image segmentation, based on an encoder-decoder design with linear complexity. Evaluated on CT pancreatic segmentation datasets for cancer diagnosis, our approach achieved a 91.32% accuracy with fewer annotations, ensuring practical segmentation performance and precise uncertainty quantification.
Dec 2022: Awarded 1st Prize worth 2500 USD[] among 1000 teams nationwide by the Chief Minister of Odisha at 2022 Smart Odisha Hackathon .
Aug 2022: Ranked 2 out of 100 participating teams at the New York City Gradio Hackathon [] and won 200$ USD worth of Hugging Face store goodies
July 2022: Received a Full-Ride grant to attend 2022 DLMI Summer School (3rd Edition) [] at ÉTS-Montreal.
May 2021: Nominated for the OUTR Best Thesis Award for the batch of 2021 consisting of over 1200 students.
Dec 2020 & May 2021: Awarded OUTR Merit Scholarship for 1st rank in the department and among top 5% of the 2021 batch for the last two undergraduate years.