I am a graduate student pursuing my Master's in Electrical and Computer Engineering at Carnegie Mellon University. I have completed my Bachelor's in Electrical and Electronics Engineering with a minor in Computer Science from the National Institute of Technology, Tiruchirappalli in India.
I have a knack for problem-solving paralleled with my dedication towards making a positive impact on society, which motivates me to contribute to the ever-evolving landscape of technology and engineering.
Outside academia, I have been a long-standing member of Toastmasters. Being a part of the club as a member, vice president and then president has helped nurture me into an effective communicator and leader. I have been a part of NGOs like Children of the World, and also the Lockdown Relief Project in India. On a personal note, I like working out and playing sports.
I am always interested in collaborating on projects and have experienced a diverse blend of industry and academic research encounters.
Portfolio
Timeline
Machine Learning Engineer Intern @ HP
Machine Learning Engineer Intern @ Ayar Labs
Started Master's in Electrical and Computer Engineering @ CMU
Deep Learning Engineer Intern @ Skylark Drones
Graduated with Honors from NIT Tiruchirappalli
Research Intern @ Hubert Curien Lab under Prof. Hubert Konik
Research Intern @ Indian Statistical Institute (ISI) under Prof Umapada Pal and Tapabrata Chakraborty PhD
Software Engineering Intern @ Raptee Energy
Research Intern @ Indian Institute of Technology (IIT) BHU, Under Prof Rajeev Srivastava
Started undergrad in Electrical and Electronics Engineering at NIT Tiruchirappalli
Professional Experiences
Machine Learning Engineer Intern @ HP Inc
Working on model optimizations using pruning to deploy object detection models on edge-devices.
Implementing unstructured and structured pruning techniques to induce sparsity into the model without performance drop.
Using pruning for model size reduction and inference speedup.
Machine Learning Engineer Intern @ Ayar Labs
Developed end-to-end machine learning workflow for automating visual inspection in optical I/O chiplet
manufacturing using vision transformers. Achieved 90% accuracy in automated visual inspection, resulting in a 6x increase in manufacturing
efficiency and significant resource savings. Utilized computer vision and optical character recognition (OCR) to accurately detect die location and
number on wafers, optimizing quality control processes.
Deep Learning Engineer Intern @ Skylark Drones
Actively contributed to the advancement of AI-powered analytics on Spectra, a drone data platform,
enabling visualization of worksites and supporting sustainable project planning.
Spearheaded and optimized machine learning pipelines for road and haul road detection, collaborating
with cross-functional teams to seamlessly integrate AI analytics within the Spectra platform. Achieved
an Intersection over Union (IoU) score of 90% on road segmentation tasks.
Pioneered innovative approaches for crest and toe detection on mining sites by leveraging edge
detection and segmentation techniques. Implemented these novel methods to enhance site analysis and decision-making processes.
Software Engineering Intern @ Raptee Energy
Built and orchestrated the backend infrastructure for Fossil Dashboard, an interactive platform for users to track live automobile data, on AWS
Implemented serverless microservices using Lambda functions, S3, CloudWatch, API Gateway, and DynamoDB to ensure seamless data processing and storage.
Configured IAM roles, facilitating secure access and permissions management for user-managed roles and successfully integrated backend services with the frontend team.
Research Experiences
Research Intern @ Indian Statistical Institute (ISI), Kolkata
Under Prof Umapada Pal and Tapabrata Chakraborty PhD
Conducted innovative research in the field of medical imaging, focusing on the application of Explainable Artificial Intelligence (XAI) to enhance fairness and transparency in skin cancer classification.
Contributed in developing a comprehensive framework to compare the features learned by deep neural networks against clinical features utilized by dermatologists. Employed qualitative analysis with class activation mappings and quantitative evaluation using dice scores.
Successfully published and presented the research findings at the Medical Image Understanding and Analysis (MIUA) conference held at the University of Oxford.
Research Intern @ Universite Jean Monnet, Saint-Etienne France
Under Prof Hubert Konik at Hubert Curien Lab
Conducted a comprehensive study of computer vision and deep learning methods for tree classification based on bark images.
Evaluated and compared the performance of various computer vision techniques, including Local Binary Pattern (LBP), Gray Level Co-Occurrence Matrix (GLCM), Gabor filters, Scale-Invariant Feature Transform (SIFT), and Histogram of Oriented Gradients (HOG) for feature extraction. These techniques were utilized to classify tree species accurately from bark images.
Implemented transfer learning using the ResNet-18 architecture on the Trunk12 Dataset, which consisted of only 393 images. Achieved an accuracy of 70% in tree species classification, building the foundation work for further advancements in tree classification and monitoring.
Research Intern @ Indian Institute of Technology (IIT) BHU, Varanasi
Under Prof Rajeev Srivastava
Implemented the U-Net architecture for medical image segmentation, specifically targeting brain images. Conducted experiments to assess the performance of the M-Net and augmented the traditional U-Net with skip connections. Achieved a dice score through model evaluation and performance optimization.
Personal Projects
Catastrophic Forgetting in LLMs for QA, summarization and sentiment analysis
Focuses on developing an accurate forecasting model for the Air Quality Index (AQI) in Pittsburgh, specifically targeting particle pollution (PM2.5) and Ozone concentrations.
The project employs a systematic approach involving data preparation, interpolation, stationarity and correlation tests, and model development. Baseline models, including the Classification and Regression Tree (CART) algorithm and Autoregressive Integrated Moving Average (ARIMA), are enhanced by integrating additional weather and pollutant data.
Experimental results demonstrate that the enhanced XGBoost model outperforms others in terms of Root Mean Square Error (RMSE).
Bird’s Eye View(BEV) in Bad Weather Conditions
Developed a multi-sensor fusion system integrating camera, LiDAR, and radar data for robust BEV(Bird’s Eye View) object detection in adverse weather conditions.
Implemented and optimized TransWeather model for real-time image preprocessing to enhance visual data quality in rain, snow, and fog.
Conducted comparative analysis of sensor fusion techniques for object detection using Detectron2 on RADIATE dataset.
Solving data scarcity in medical imaging using Diffusion models
Implemented and fine-tuned diffusion models to generate synthetic skin lesion images, addressing data scarcity in medical imaging.
Developed a conditional image generation pipeline using Stable Diffusion with LoRA achieving high-quality results with only 700 training images.
Conducted comparative analysis of unconditional and conditional image generation techniques using various data subset sizes and evaluated performance on improving downstream classification performance on skin lesion types using Vision Transformers and ResNet architectures
Fine Grained Image generation using Attention GANs
Enhanced text-to-image generation with fine-grained semantic control using advanced neural networks.
Improved the AttnGAN architecture by integrating BERT and CLIP models for more accurate and detailed image outputs. Conducted ablation studies to validate the performance improvements of the modified model on MS COCO dataset.
Time Series Analysis for Air Pollution Data
Focuses on developing an accurate forecasting model for the Air Quality Index (AQI) in Pittsburgh, specifically targeting particle pollution (PM2.5) and Ozone concentrations.
The project employs a systematic approach involving data preparation, interpolation, stationarity and correlation tests, and model development. Baseline models, including the Classification and Regression Tree (CART) algorithm and Autoregressive Integrated Moving Average (ARIMA), are enhanced by integrating additional weather and pollutant data.
Experimental results demonstrate that the enhanced XGBoost model outperforms others in terms of Root Mean Square Error (RMSE).
End-End Question Generation and Answering
Developing end-to-end pipelines for Question Generation (QG) and Question Answering (QA) tasks using transformer-based models. The work emphasizes the use of transformers, specifically the T5 model for QG and DistilBERT for QA.
The QG model aims to generate diverse and non-deterministic questions by leveraging transformer capabilities. The QA model is trained on the SQuAD dataset using DistilBERT and addresses challenges related to handling long contexts and answering polar questions.
Experimental results and hyperparameter tuning are discussed for both QG and QA models, highlighting the importance of specific configurations in improving model performance
Geometric 3D Reconstruction
Implemented a robust algorithm involving feature detection, camera pose estimation, depth estimation, surface reconstruction, and texturing, showcasing expertise in computer vision and image analysis.
Enhanced the reconstruction model by incorporating LOFTR features using deep learning, significantly increasing the density of 3D representations and addressing challenges of sparsity.
Implemented Dense Matching techniques, expanding the project's capabilities by providing additional correspondences beyond sparse keypoint matching.
Cyber Attack Detection in Power System SCADA Networks using Machine Learning Techniques
Final year undergraduate project to simulate faults and cyber attacks like man in the middle and analyze real time data flow in SCADA networks. Applied machine learning techniques like K-Means clustering, Decision Tree classifier and Random forest to distinguish cyber attacks from naturally occurring events.
Promoting Innovation in National Institute Of Technology Tiruchirappalli
Hackathon project to simplify the process of internship, project collaboration and placements by creating a unique Research Quotient for every candidate. Implemented in at NIT Trichy and selected for Smart India Hackathon 2020.
EXTRACT - Text Extraction
Inspired from Google Tesseract and based on Convolution Neural Networks as well as OpenCV. EXTRACT is an Optical Character Recognition (OCR) engine to extract text from images and convert them to plain text. (Dec ’20)
(detail more)
Image Background Remover
Using Computer Vision and Deep Learning to remove background of input image.
(detail more)
Air Pollution Analysis
Analyzed air pollution data in New Delhi over the past 20 years by studying specific pollutant levels in each area and visualized the pollution patterns using Python, Pandas, NumPy,Matplotlib and Seaborn libraries.
Controller
Locked Safe
Lorem ipsum dolor sit amet, consectetur adipisicing elit. Mollitia neque assumenda ipsam nihil, molestias magnam, recusandae quos quis inventore quisquam velit asperiores, vitae? Reprehenderit soluta, eos quod consequuntur itaque. Nam.
Submarine
Lorem ipsum dolor sit amet, consectetur adipisicing elit. Mollitia neque assumenda ipsam nihil, molestias magnam, recusandae quos quis inventore quisquam velit asperiores, vitae? Reprehenderit soluta, eos quod consequuntur itaque. Nam.