Data Scientist
Jaggaer-Docskif
Total des années d'expérience :8 years, 0 Mois
• Building POCs for Q&A and entity extractions on contract documents using Lang-chain and POCs on Clause Com parison and scoring along with reasoning using LLMs on AWS Bedrock. Also fine tuned Mistral-7B model for entity extractions on contract documents.
• Implemented end-to-end named entity extraction (NER) pipelines to train and deploy deep learning models, which enabled annotators to upload contract PDFs, label the entities, independently train and infer or extract data without the need for Data Science experts.
• Implemented NER labelled data validation using confident-learning techniques, resulting in 70% reduction in the re viewing time of labelled data, also significantly improved data quality for training.
• Built an eficient Document Title Extraction model with an F1-score of 90% and parallelly reduced latency to 300ms per PDF document using features transformed from OCR outputs, previously title extraction relied on object detection models.
• Fine-tuned YOLO object detection models with custom data augmentation for signature detection from PDFs, achieved an F1-score of 85%, which led to substantial improvement in document compliance eficiency.
• Developed user-friendly APIs for the insurance industry to extract essential details, such as meta-data and organ abnormalities, from pathology reports, using OCR, regex, and custom NER techniques.
• Designed and implemented a chatbot for online fee payment applications, including data preparation tasks such as intent and entity annotation. Trained the chatbot using the RASA framework.
• Developed a Face Recognition biometric system for a sports academy on Nvidia Jetson Nano.
• Managed the migration of databases from Oracle to PostgreSQL for TCS Alumni Portal and Ultimatix Gems, ensuring a smooth transition. Collaborated with cross-functional teams for a successful migration, enhancing data management eficiency and system reliability.
- Gait analysis of Biped robot, simulated in MS Adams