Projects
Research papers, production pipelines, and side experiments — things I've built that I'm proud of.
Featured Projects
TinyYoloV2 Face Detection Visualizer
Built a live, interactive browser-based tool that visualizes how a face detection AI "sees" the world — layer by layer, neuron by neuron. Most AI models are black boxes; you put in an image and get an answer but have no idea what happened in between. This project peels back those layers, letting users literally see what patterns each part of the neural network is detecting. Built during a research collaboration with Carnegie Mellon University (CMU), this tool was built using TensorFlow.js, enabling it to run entirely in a web browser with no installation required.
Aspect-Sentiment Opinion Summarizer (Published Research)
Built an AI system that reads e-commerce product reviews, Q&As, and descriptions from multiple sources and produces a structured, aspect-based summary — telling you exactly what customers love or hate about specific features like battery life, design, or price. The system scraped real product data from Flipkart and used generative AI models (T5, mT5, GPT-3) to generate human-readable summaries. This research was published at ACM CODS-COMAD 2023, one of India's premier data science conferences.
Docketry — Intelligent Document Processing Platform
Designed and built the AI core of Docketry, a SaaS platform that automates document-heavy business workflows. The platform can read any document — invoices, contracts, ID cards — understand its layout and content, and automatically sort it into the right category. The AI model I fine-tuned, LayoutLMV2, reads documents the way humans do: it looks at both the text AND the visual structure of the page simultaneously. I also built the entire server backend, designed the system architecture, and published a Python package so businesses could plug Docketry into their existing systems with just a few lines of code.
Indic NER — Multilingual AI for Indian Languages
Built an AI system that can read mixed-language text — like when someone writes a sentence using both Hindi and English words — and automatically identify important entities such as movie names, actor names, and locations. This is called Named Entity Recognition (NER), and doing it for Indian languages is uniquely hard because people switch languages mid-sentence all the time. I built this for the top 5 Indian languages, covering 17 different entity types, trained on over 100,000 annotated entries. I also created a data augmentation technique to generate more training data for languages with limited resources, achieving a 20% improvement in accuracy.
RAG Pipeline for Healthcare Intelligence
Built an AI-powered system that makes sense of massive healthcare databases in real time. Imagine asking a question in plain English and getting back a concise, accurate summary from thousands of patient records — that is what this system does. It uses a technique called Retrieval-Augmented Generation (RAG), where an AI first searches for the most relevant records, ranks them by how closely they match the question, and then generates a clean, readable summary. The system was deployed in production with parallel processing, handling multiple queries simultaneously without slowdowns.
All Projects
Scroll horizontally to explore