- Part of development team of one of the most prominent NLP libraries: Spark NLP.- Lead developer of NLPTest library that tests large and traditional language models on more than 50 different criteria including biasness and robustness.- Helping secure multiple six figure enterprise customers by developing POCs and then working on long-term projects to bring them to fruition.- Defining project roadmap for both LangTest and SparkNLP libraries: implementing, testing, andreleasing new features.- Working with sales team to secure licensing agreements with prospects- Writing and contributing to open-source tutorial notebooks, blogposts, webinars and documentation for the community- Research and development of ML models for bespoke healthcare/clinical use-cases: - Reduced large language model size by performing distillation for healthcare domain. This resulted in a higher compression ration, reducing the model size by 50% while maintaining similar level of performance, which helped large scale deployment of the models. - Improved contextual factuality of the models using augmented data training to reduce biasness. - Research on adversarial evaluation of LLMs to improve robustness.- Implementation and improvement of Q&A (RAG) engine(s): - Implemented various ranking improvement techniques (like reranking during inference and smart chunking and metadata integration during generation etc). Improvements in ranking alone improved MAP by 20 points. - Tuned models (DPO) especially for RAG to improve the final result while reducing model size to improve scalability.- Collaborating with team members and customers to publish over 1000 different models for various use-cases including Summarization, Fact-checker, Named Entity Recognition, Entity Disambiguation, Entity Linking, Assertion detection, etc.- Working on MLOps to design model life cycle management including training, deployment, monitoring, and drift detection.

Background remover:Developed a image background model for natural images using object segmentation (deeplab / detectron, U2) models for background segregation, and image matting models to get smooth edges.Image enhancer:Developed a GAN model - trained on synthetic data generated by adding noise and reducing size of baseimages – capable of removing noise & upscaling images.Fraud Detection:Developed a complete fraud detection system by utilizing billions of rows to find fraudulent patterns in mobile financial systems using statistical and ML models. Leveraged BigData technologies to handle large amounts of data, and PySpark for analysis & modeling.

> Developed a data entry automation engine that automatically processed incoming scanned documents, performed classification, text localization and OCR, pre-filling the data entry forms, and reducing 80% manual data entry time (as end-user only has to perform validation and not complete data entry). The following steps were involved: - Document Image Classification. - Document Image Orientation detection. - Synthetic Data generation & Text Localisation. - OCR on text images. - Selection region detecting using object detection (Yolo).> Built a revenue forecasting model that predicted estimated revenue for each claim submitted to the insurance companies. This model identified dormant claims worth > 1M, and successful follow-up on those claims resulted in an additional revenue of > 1M.> Built a symptom checker app that asked patients questions based on entropy, and honed down to ICD 10 codes.- Clustering of textual laboratory data using NLP.- Integration of AI in MTBC's EHR (TalkEHR); Automated the coding process by suggesting Procedure Codes to doctors, drastically reducing the charting time.- Diagnose code prediction using advanced data science algorithms- Prediction of health insurance claim payments and aging (MTBC Same Day Funding)