CV

Master of Science

Northeastern University

Artificial Intelligence

CGPA : 3.76

• Algorithms, Foundations of AI, Unsupervised ML, Pattern Recognition and Computer Vision, Machine Learning, Mobile Robotics, Verifiable ML, AI for HCI

Computer Vision Intern

Norfolk Southern

Developed an agentic RAG solution and engineered GPU-accelerated, high-throughput computer vision and Big Data pipelines.

• Built a semi-automatic labeler that proposes label priors, which was shown to 3x human semantic segmentation labeling speed.
• Integrated an automated data-augmentation prompter that generated polylines to guide, accelerate, and automate data curation, seamlessly importable into the CVAT platform for training defect detection models.
• Developed a PyTorch-based zero-shot segmentation model that leverages depth maps from a foundational depth estimator toidentify parts of railway cars by their relative depth in the scene
• Prototyped a few-shot segmentation pipeline using VRP-SAM to accurately segment novel classes from as few as 20–50 exemplars, compared to conventional models that require 5,000–10,000 labeled images for fine-tuning.

Generative AI Intern

Inflohealth

Built a language model pipeline that processes millions of radiology reports to curate datasets for downstream healthcare analytics.

• Implemented a multimodal RAG workflow on the Qdrant vector database for large-scale PDF-based question answering.
• Developed a large-scale PDF retrieval pipeline using the ColPali model, improving top-5 recall by 8% over OCR solutions.
• Built a multi-agent system using LangGraph that divides the task into sub-tasks, leverages a document retriever, a metadata crawler, and feedback loops to maximize cited content in the generated answer.
• Implemented a system prompt refinement module using Mutual Information Maximization on LLM-generated candidates which demonstrated an average reduction of 6.2 steps for task completion, and a 14% increase in task completion rate.
• Built a multi stage reranking based RAG workflow with LlamaIndex for retrieval based on medical rules to maximize.
• Migrated the data loader pipeline to NVIDIA DALI and Ray, boosting throughput of large-scale model evaluation by 2.1 times.

Artificial Intelligence Intern

Sentient.io

Customized open-source computer vision models with application-specific optimizations for AI microservices.

• Implemented a video action-recognition pipeline, quantizing the model with TensorRT to reduce latency by 12%.
• Reduced video object detection misclassification rate by 7% using a Kalman filter motion model.
• Built a general-purpose auto-labeler using SAM and GroundingDINO, which was applied to three datasets (15,000+ images).

Artificial Intelligence Engineer Intern

Juhomi

Co-developed an AI-powered, microservice-based retail analytics platform from the ground up; it was contracted by four MNCs.

• Managed a crowdsourced data annotation job for object detection of 252 product classes across 4750 images in Amazon Mechanical Turk, and developed a YOLOv5-based auto-labeller in Amazon SageMaker to extend the dataset to 21,000 images.
• Integrated image super-resolution and optical character recognition (OCR) improving mAP by 0.2 over naive object detection.
• Demonstrated the combined detection pipeline’s ability to perform zero-shot object detection of previously unseen classes.
• Upgraded the API architecture to support asynchronous communication using FastAPI with Uvicorn, enabling an API throughput increase of 341% (i.e., from 12 requests per second to 53 requests per second).
• Developed continuous training and monitoring workflows using Apache Airflow, Data Version Control, and MLflow, enabling class-specific model updates through model-guided data selection (active learning).

HuskyGuide: Database Interface Agent (sponsored)

• Spearheaded the development of a multi-agent system leveraging Northeastern University’s knowledge base for policy navigation.
• Developed a SQL agent that generates complex queries through multi-hop reasoning using execution plans and database contents.
• Illustrated the algorithm’s behavior and efficacy through comprehensive visualizations

ProdSeek: Semantic Product Catalog Search

• Developed a recommender that takes product selections from images and suggests similar products from a product image corpus
• Built an adaptive vector search using YOLOv3; embeds live product selections and performs real-time semantic retrieval.
• Conducted ablation studies demonstrating higher product specificity and qualitative semantic capacity than ResNet embeddings.

RobAnn: Exploring Neural Network Robustness

• Developed an algorithm to quantify deep neural networks’ resistance to adversarial attacks and noisy perturbations.
• Performed feature-wise robustness analysis to expose adversarially susceptible neurons by targeted weight perturbation.
• Implemented it with LangChain and packaged it as a FastAPI app that supports integration into concurrent multi-agent systems.

AR-enabled textbooks

IEEE

• Published in 2023 4th International Conference on Electronics and Sustainable Communication Systems (ICESC)
• Built an Augmented Reality mobile application that scans QR codes embedded in textbooks to render animated 3D models.
• Secured TNSCST funding to deploy the AR pipeline in 9th-grade textbooks for the Tamil Nadu state curriculum.

Transformers for Browse Node Classification with Class Imbalance

IEEE

• Published in the 2023 International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES)
• Fine-tuned the DeBERTa transformer model for e-commerce classification on 10M+ records across 250 browse node categories
• Applied Focal Loss to mitigate severe class imbalance by down-weighting the gradient updates of dominant classes over time.
• Achieved an increase of 2% in validation accuracy with faster convergence compared to vanilla DeBERTa and other BERT variants.

Safety surveillance using Explainable Object Detection

Springer

• Published in Smart Trends in Computing and Communications (SmartCom 2023) as a part of the book series: Lecture Notes in Networks and Systems ((LNNS,volume 650))
• Built an automatic object detection explainer that visualizes the top three salient activation regions influencing predictions.
• Implemented a novel model-agnostic pipeline that automatically finds salient layers to bypass manual exploration.
• Demonstrated the pipeline on an artificially biased dataset simulating effects of irresponsible data collection practices.
• Augmented the object detection pipeline with Sobel features to improve generalizability and reduce bias.