Building production-grade AI assistants and retrieval systems that blend multimodal search, orchestration, and observability.

• Designed a multimodal RAG pipeline using CLIP embeddings and ChromaDB to retrieve text, images, audio, and video frames, pairing them with GPT-4o for context-aware answers and tuned latency/accuracy trade-offs.
• Engineered a Swarm-driven multi-agent router with OpenAI function calling, dynamic tool registry, Pinecone-backed knowledge bases, and Django/DRF persistence plus analytics for chat health.
• Built an AI workflow marketplace with a unified LLM layer (OpenAI/Anthropic/Llama), multi-tenant API management, and LangGraph agent templates with evaluation harnesses.

Domain: GenAI, Multi-Agent Systems, Retrieval-Augmented Generation, Backend Engineering

Enhancing student learning with LLMs that understand long educational videos end-to-end.

• Built a data pipeline combining manual annotation and Gemini 1.5 Flash synthesis to create 4,300+ high-quality Q&A pairs across lecture domains.
• Benchmarked multimodal LLMs (Video-LLaMA, mPLUG-Owl, etc.) across zero/few-shot and fine-tuned settings, evaluating BLEU, ROUGE, METEOR, and qualitative student feedback.
• Co-authored the EMNLP 2025 paper on educational multimodal video question answering.

Domain: Multimodal AI, Natural Language Processing, Education Technology

Self-supervised pipelines for robust 3D Cryo-ET analysis in open-set environments.

• Integrated Momentum Contrast with ViVIT/ViT (inflated weights) for 3D Cryo-ET classification and improved representation quality.
• Designed a 3D processing workflow that boosted performance and training efficiency, achieving 70.28 F1 in open-set settings.
• Contributed experiments and writing for an ICCV 2025 manuscript and a book chapter on feature semantic segmentation.

Domain: Computer Vision, Self-Supervised Learning, Computational Biology

The main focus of this project is the development of IndicBART, a multimodal summarization framework that integrates text and image data to produce regionally contextualized summaries in Indian languages. The process involves the following steps:

• Collected a multimodal dataset (Hindi, Tamil, Bengali, Marathi) with paragraphs, summaries, and images from Large Scale Multi-Lingual Multi-Modal Summarization Dataset (M3LS).
• Extended the BART model with a visual-aware encoder to generate regionally contextualized summaries using both text and image data.
• Achieved a ROUGE-1 score of 0.266 on the Hindi dataset and 44.1% image precision using an image pointer for multi-output prediction.
• Work accepted at ICDAR 2024 based on multimodal summarization research with Indian regional languages.

Domain: Natural Language Processing, Multimodal AI, Summarization, Machine Learning

This project aims to develop an effective pneumonia detection system by combining advanced image analysis techniques, achieving high accuracy in classification. The workflow encompasses the following steps:

• Compiled and preprocessed a Mendeley dataset with three classes (Covid-19, Pneumonia, Normal), applying data augmentation techniques to address overfitting.
• Extracted image features using Vision Transformer (ViT) and ResNet50 models.
• Developed an ensemble model by combining features from both ViT and ResNet50, and used XGBoost for classification, achieving 96.19% accuracy.
• Contributed to writing the results section of the paper, which was accepted at CPAMCS-2023.

Domain: Medical Imaging, Computer Vision, Machine Learning, Image Classification

This project develops a multimodal framework that combines text and image data to detect complaints in customer reviews. The process involves the following steps:

• Collected a multimodal dataset of product images, customer reviews, and aspects from Amazon website.
• Utilized BERT for text data, and VGG16/ResNet models for visual embedding.
• Developed a multimodal interaction model to learn the relationship between text, image, and product aspects.
• Identified complaint eligibility by analyzing multimodal relationships, with the paper published in ECIR 2023.

Domain: Natural Language Processing, Multimodal AI, Sentiment Analysis, Image Processing

Predicting vessel paths and collision risk from noisy AIS streams.

• Cleaned AIS data (SOG, COG, MMSI, etc.) and clustered vessel patterns with DBSCAN/K-Means to remove outliers.
• Built sequence-to-sequence baselines (RNN, LSTM, GRU variants) for trajectory forecasting and risk assessment.

Domain: Time Series Modeling, Machine Learning, Maritime Analytics

Optimizing concrete strength and durability through data-driven mix modeling.

• Experimented with sand fines, cement, and plasticizer ratios, comparing regression models (MLR, SVR, Decision Trees, ANN) for strength prediction.
• Achieved top accuracy for 7-day and 28-day compressive strength using Decision Tree Regression.

Domain: Applied Machine Learning, Regression Modeling, Materials Engineering