Deepak Sahu

Led the development and production deployment of a GenAI claims intelligence platform on Databricks for an Australian insurer, automating home claims document analysis through summary generation, expert report synthesis, evidence-based citations, and damage classification. The solution achieved 95.07% completeness and 99.70% relevance, driving substantial reviewer productivity gains and protecting more than $15M in at-risk business value. Designed GTM use-cases for the NVIDIA's Metropolis framework.

Reduced cost to adoption of partner's stack by 73%. Developed multi-use agentic architecture, SAAS and Chat, using NVIDIA Nemo Agent Toolkit and LangGraph.

Senior Data Scientist at EXL Services (2025-04 – 2026-07)

• Led the development and production deployment of a GenAI claims intelligence platform on Databricks for an Australian insurer, automating home claims document analysis through summary generation, expert report synthesis, evidence-based citations, and damage classification. The solution achieved 95.07% completeness and 99.70% relevance, driving substantial reviewer productivity gains and protecting more than $15M in at-risk business value. • Designed GTM use-cases for the NVIDIA's Metropolis framework. Reduced cost to adoption of partner's stack by 73%. • Developed multi-use agentic architecture, SAAS and Chat, using NVIDIA Nemo Agent Toolkit and LangGraph.

Gen AI Senior Associate Consultant at EY Global Delivery Services India LLP (2024-05 – 2025-04)

• Created multi-agent conversation agents with instruction tuning Large Language Model LLM (GPT-3.5 turbo) and tool calling to create deterministic flow to access databases, formulate and execute complex SQL queries from natural language query; architecture vertically scaled the context-window length of LLM. • Consulted teams to implement hybrid search (BM25 + Similarity Search) in conjunction with hypothetical document embedding technique (HyDE) for better retrieval of cross-referenced terms in tax rules; increased coverage of stakeholder questions by 40%

Data Science Associate Consultant at SAP India Pvt. Ltd (2024-01 – 2024-05)

• Implemented RAG application to augment tasks of Virtual Customer Success Partners (VCSP) up-to-date knowledge and generate personalized response; Improved VSCP productivity by 15% • Researched, Designed Coded and Deployed customized Python packages to evaluate different chunking algorithms (fixed-size vs semantic) and libraries (Pinecone's hybrid search vs ChromaDB 's cosine) to be used in RAG framework

Data Science Scholar at SAP Labs (2021-08 – 2023-12)

• Trained NER using Spacy and built a few-shot based framework to embed Knowledge graph triples to natural queries; achieved a comparable result to a fine-tuned model with ±2% on (1-shot & 2-shot) queries. • Designed mathematical model to optimizing large-scale data transformation pipelines for improving analytics in supply chain visibility of Oil & Gas sector; reduced time of report generation by 87% • Collaborated with cross-functional teams and authored technical implementation for ML Ops in SAP AI Core.; Contributed in adoption of product by 25 clients.

Python Backend Developer at Bipolar Factor (2021-03 – 2021-07)

• Designed and deployed face-recognition based door access mechanism for manufacturing industry floor. Ground application built in python for Raspberry Pi 3 (Linux, raspbian) and cloud logging on AWS (Rekognition, Lambda, S3).