Priyatosh Raut

Data Scientist with 6+ years of experience in NLP, machine learning, and end-to-end ML system design across fintech and healthcare. Proven track record of designing scalable ML pipelines, reducing operational cycle times by 85%, and delivering measurable impact through fraud detection, feature engineering, and predictive modeling. Experienced in owning the full model lifecycle, authoring system design documents, and mentoring junior engineers.

Data Scientist 2 at Setu By PineLabs (2025-01 – Present)

• Developed core model layers for SignalIQ (Bank Statement Analyser) — heuristics-based and multi-class ML classification pipelines.
• Built an end-to-end Annotation Pipeline for cold-start transaction categorization using FAISS, multi-stage deduplication, intelligent sampling, and a 3-model validation layer (ChatGPT, fine-tuned LLM, P2P BERT). Reduced POC turnaround from ~3 months to ~2 weeks; scaled golden dataset from 2L to 30L with 80% less annotation.
• Designed Borrower Tags framework — interpretable behavioral features (affluence, spend, debt, credit behavior) — improving defaulter identification by +4% AUC.
• Fine-tuned Gemma 3 for NER, extracting 5 entity types (counterparty, merchant, transaction metadata) from bank narrations.
• Built Salary Detection model using BERT and Jaccard similarity with more than 90% precision and recall for salary and earning-related-to-salary identification.
• Led 6+ POC deliveries end-to-end within 2-week turnaround, achieving more than 95% precision and recall.
• Designed and developed in-house PDF parser using OCR, PDF extractors, and LLMs (prompt engineering) to replace external vendor; achieved ~90% feature parity via systematic A/B testing and experiment design.
• Authored 6+ design documents for ML system architecture; mentored 3 junior DS and integrated AI-assisted tools (Claude Code, Codex) for rapid prototyping and accelerated delivery.

Data Scientist at Medi Assist Healthcare Services (2022-02 – 2024-12)

• Increased fraud detection accuracy by 30% using ensemble ML algorithms (XGBoost, Random Forest, Isolation Forest) on healthcare claims data.
• Developed an entity detection model to streamline healthcare claim processing, improving operational efficiency by 25%.
• Implemented an in-house NER-based account verification system, reducing third-party costs while improving accuracy and latency.
• Built and deployed scalable REST APIs using Flask and FastAPI to integrate ML models with production systems.
• Applied LSTMs, XGBoost, and Isolation Forest to optimize product development and anomaly detection.
• Optimized ML workflows using MLOps tools (MLFlow, Kubernetes, Docker, Grafana, New Relic) for scalable model deployment.
• Led a team of interns to enhance existing healthcare analytics products.

Machine Learning Engineer at SureClaim (merged with Connect and Heal Primary Care) (2020-01 – 2022-02)

• Conducted POCs to evaluate potential solutions and successfully deployed the best options using Kubernetes.
• Developed NLP models using AWS Textract and LSTM to automate health insurance claim processing, including text classification and named entity recognition.
• Built predictive models using XGBoost, Random Forests, and Bayesian Networks for various applications.
• Managed the complete lifecycle of model development, from data collection and cleaning to deployment and ongoing maintenance.

Machine Learning Intern at Brigosha Technologies (2019-06 – 2019-07)

• Developed a proof of concept using CNNs and image processing to automate amount reimbursement from receipts.
• Achieved over 70% accuracy in text recognition by building an incremental CNN model for OCR on financial documents.

B.Tech in Computer Science – Tezpur University (2016 – 2020)