Professional Journey

Building AI-powered healthcare software from Berlin, Germany. My work spans two worlds: the front-end (what users see and click on) and the back-end (the invisible engine that makes everything run). On the front-end, I improved the application's interface using Node.js to make it more intuitive and responsive for end users. On the back-end, I am leading the redesign of the product's core architecture — essentially re-engineering the "nervous system" of the application to make it handle more requests faster, with less waiting time. This is achieved using FastAPI, a modern Python framework known for its exceptional speed and efficiency.

• Optimized the user-facing interface in Node.js, enhancing usability and engagement for healthcare platform users
• Leading architecture design using a FastAPI hybrid framework to significantly reduce latency and improve platform performance
• Applying Gen AI and Natural Language Processing techniques to solve real-world challenges in the healthcare domain

Worked at a leading analytics consultancy, building intelligent data systems for healthcare clients using the latest AI technologies. My most significant contribution was designing a RAG (Retrieval-Augmented Generation) pipeline — think of it as an AI-powered research assistant that not only searches through massive databases to find the most relevant records, but also reads and summarizes them in plain language, ranked by relevance. I also built the full production infrastructure to deploy this system reliably at scale. Beyond AI, I automated data collection by building web crawlers — programs that automatically browse websites and extract useful information — cutting manual effort by 85%. I also developed tools to read unstructured documents (such as reports and articles) and automatically fill in missing fields in master datasets.

• Built a RAG pipeline using Large Language Models (LLMs) to generate intelligent query filters and summarize healthcare records, with accuracy improved through similarity-based ranking
• Engineered an end-to-end production inference pipeline optimized with parallel processing, significantly reducing execution time at scale
• Designed automated web crawlers using Selenium and async programming (asyncio, aiohttp), improving data pipeline efficiency by 85%
• Extracted key entities from unstructured documents using LLM prompt tuning for classification, enriching master datasets and enabling downstream analytics

Led AI and automation development for Docketry — an intelligent document processing platform that helps businesses automate paperwork workflows. My core contribution was fine-tuning a state-of-the-art AI model called LayoutLMV2, which reads documents the same way humans do: by understanding both the text and the visual layout of the page. I trained this model to classify documents into 10 different categories with 94% accuracy. Beyond the AI model, I designed and built the entire server infrastructure — including the API, authentication system, and cloud deployment — and published a Python package so clients could integrate Docketry into their own systems with minimal effort.

• Fine-tuned the LayoutLMV2 document AI model, achieving 94% F1-score across 10 document categories — enabling automated, highly accurate document classification
• Built and led development of a gRPC-based Django server with full authentication and authorization, deployed on Azure with Nginx for SSL-secured production-ready hosting
• Published the Docketry PyPI package to enable seamless, plug-and-play client-side integrations across diverse business systems
• Explored prompt-engineering techniques to extract structured fields from raw text documents, improving processing speed and accuracy for automation workflows
• Designed user analytics dashboards to visualize usage logs and provide meaningful summaries in the Docketry dashboard

Conducted advanced research in Natural Language Processing (NLP) — the branch of AI that enables computers to understand and process human language — under Prof. Pushpak Bhattacharya, one of India's foremost NLP experts. My research focused on a uniquely Indian challenge: understanding code-mixed text, where people blend two languages in a single sentence (e.g., mixing Hindi and English). I built a Named Entity Recognition (NER) system — an AI that automatically identifies names of people, places, movies, and other entities in text — for the top 5 Indian languages, covering 17 entity types and trained on 100,000+ data entries. I also developed data augmentation techniques to create training data for low-resource Indian languages, leading to a 20% performance improvement.

• Built a multilingual NER model for code-mixed Indian language queries covering 17 entity tags and 5 Indian languages; fine-tuned an ensemble BERT model achieving 94.79% F1-score
• Developed a data augmentation pipeline using open-source tools and a custom gazetteer list, achieving a 20% improvement in NER performance for targeted low-resource language tags
• Explored a wide range of deep learning architectures — RNNs, LSTMs, Transformers, T5, and GPTs — across NER, transliteration, machine translation, and summarization tasks
• Published research paper: 'Aspect-Sentiment-based Opinion Summarization using Multiple Information Sources' at ACM CODS-COMAD 2023