> whoami
Designing and building end-to-end AI systems, with core expertise spanning predictive analytics, Computer Vision, NLP, and advanced generative models like LLMs and VLMs.
Built a custom RAG pipeline from scratch including ingestion, chunking, vector DB and integrating local FlashRank reranking to maximize retrieval accuracy without added API costs. Engineered a non-blocking architecture with isolated FAISS stores and sliding-window memory for resilient, long-running conversational queries.
Engineered a sequence-to-sequence Transformer architecture from scratch, including custom encoder, decoder, and attention mechanisms, based on the original "Attention Is All You Need" paper. Built an end-to-end training and evaluation pipeline for automated translation on a custom English-Bengali dataset.
Developed a high-performance search engine utilizing OpenAI CLIP and Qdrant vector database for both natural language and image-based retrieval. Leveraged high-dimensional vector embeddings to bypass traditional keyword search, accurately matching images based on contextual meaning.
Developed a real-time, multi-label text classification REST API utilizing custom regex preprocessing and TF-IDF vectorization to detect toxic language. Trained and tuned balanced Logistic Regression models on the Jigsaw dataset, fully containerizing the modular application for scalable deployment.
Developed dual Streamlit applications within a single repository to perform semantic word searches using BERT embeddings, allowing users to visualize semantic relationships through either word clouds or interactive network graphs.
Class Activation Maps (CAM) provide a way to identify which regions of an input image contribute most to a convolutional neural network’s decision for a specific class. CAMs highlight the spatial locations in the feature maps that are most influential for predicting class, allowing an interpretable visualization of the model’s attention. Implemented this feature for pretrained ResNet18 weights to visualize how this model actually predicts.
Class Activation Maps (CAM) provide a way to identify which regions of an input image contribute most to a convolutional neural network’s decision for a specific class. CAMs highlight the spatial locations in the feature maps that are most influential for predicting class c, allowing an interpretable visualization of the model’s attention.
Read ArticleBuild an interactive tool to explore the semantic connections inside a BERT model.
Read ArticleAccurate model validation relies on the assumption that our training and testing subsets are representative of the underlying population distribution. Stratified K-Fold addresses this by ensuring that each fold preserves the class distribution of the dataset, providing a more stable and statistically reliable estimate of a model’s generalization performance.
Read Article