Nikhil Kumar Singh Email and Phone Number

Making e-commerce search smarter.● Spearheaded the development of an image-based search feature for fashion products, leveraging transformer-based vision models (CLIP). Increased user engagement through image search, resulting in an exeptional click-through rate.● Designed and developed the data pipeline and APIs to serve thousands of users with sub 200 millisecond response time from 8 million fashion products catalog. Utilized Milvus vector database to efficiently store and retrieve image embeddings.● Leading semantic search integration to enhance search relevance. Trained language models on platform data and trends for deeper understanding. Building a lexical and semantic hybrid search for improved recommendations.● Created internal tools and pipelines with LLMs to create synthetic data for model training and improving product catalog quality. These resulted in faster model training, experimentations and increased productivity across teams.

As a Machine Learning Engineer at ConcertAI. I worked on improving end-to-end systems designed to extract Cancer Biomarkers and their multiple attributes from Oncology reports to create high precision Real World Data used for clinical trials.My everyday work revolved around leveraging language models (NLP) and OCR techniques ML and deep learning to extract entities and values from medical reports and interactions (especially Oncology).● Improved end-to-end systems designed to extract Cancer Biomarkers and their multiple attributes from Oncology reports to create high precision Real World Data used for clinical trials.● Created ML models using layout and semantic features to improve the precision of entities extracted from medical reports using annotated data. Improved precision of entity classes by 5-8%.● Optimized deep learning models using quantization and weight clustering. Helped reduce the model size and inference time by up to 34% for different models.● Minimized OCR error in alphanumeric medical entities by an ensemble of ML models.● Developed a framework to scale models and reduce inference time for processing millions of documents using a hybrid architecture with Docker, Kubernetes, and AWS Batch.● Designed a Chargrid-inspired deep learning model to understand better the layout and positioning of words in medical reports crucial for improving downstream key-value extraction.

During my time at Utopia as Data Science Engineer, I worked on creating intelligent Enterprise Asset Management solutions using ML and AI. Worked on the project from idea to production. Packaged the solution and SaaS offering for big Oil and gas and heavy machinery industries. ● Created custom word embeddings based on industry-specific data to better incorporate semantic and domain information. Used it to improve Named Entity Recognition and solve word similarity problems.● Developed and deployed an object detection model to find and localize equipment nameplates and logos from images using Faster-RCNN. Switched to a YOLO based model for better prediction time.● Built data pipelines and models to preprocess and localize text regions from images and PDFs. Used tesseract and AWS Textract for robust text extraction.● Designed and developed a real-time web scraping system to search and recommend similar mechanical equipment based on textual specifications. Leveraged NLP and business rules to match detailed specifications.● Benchmarked various OCR tools and services like tesseract, abbyy, AWS Textract, AWS Rekognition and Google Cloud Vision API for text extraction from images and documents.● Integrated machine learning solutions and pipelines. Developed and deployed REST APIs and microservices, using Django, Flask, Docker and Kubernetes.

● Worked on industry specific name mapping based on linguistic and semantic similarity of words.● Text extraction and OCR using open source libraries like tesseract and services like AWS Textract, Rekognition and Google Cloud Vision API.● Worked on intent recognition from customer complaints using Named entity recognition and tokenization.