Shobhit Maheshwari

Lead Data Scientist

Roadzen — Delhi, India

Leveraging LLMs such as Qwen to extract structured information from insurance policy documents, enabling automated policy understanding at scale.

Building an end-to-end claims underwriting agent that autonomously assesses vehicle claims and generates comprehensive inspection reports.

Research Assistant

University of Edinburgh — Edinburgh, UK

Benchmarked GPT against 4-bit quantised Llama and Mistral on FinQA, demonstrating a ~12% gain in accuracy.

Achieved a 16% improvement over RAG using a ranked MapReduce pipeline with similar processing time to RAG.

Scraped financial data sources to compile 70GB from annual reports, earnings transcripts, etc.; developing a low-latency framework to query large financial reports for all companies on a unified platform.

Senior Data Scientist

Roadzen — Delhi, India

Streamlined ML workflow from data ingestion to model serving using FastAPI, Airflow, and Jenkins, resulting in continuous training and efficient deployment of in-house ML models.

Enhanced the accuracy of the insurance policy QA bot using GPT, achieving 95% correct answer retrieval.

Developed a heuristic inspired by the concept of momentum to extract keyframes for car profiles from a video.

Achieved over 84% accuracy in classifying car colour, make, and model for expedited claims validation.

Data Scientist

Roadzen — Delhi, India

Built a Mask RCNN model for instance segmentation of damage, parts, and profile on car images using PyTorch, reducing claims processing time from 40 mins to under 2 mins.

Optimized the mAP score metric to account for subjectivity, achieving a score of 74 in damage segmentation.

Reduced Damage Recognition API turnaround time by 30% using TorchServe and FastAPI.

Earned recognition for AI models at the Asia Motor Insurance Summit and Financial Express Future Tech awards.

Data Scientist

Spoonshot — Bangalore, India

Designed a weighted DeepWalk model with ingredients as nodes and edge traversal probability based on flavour pairing theory to generate novel flavour pairings.

Managed 100 million ingredient combinations by building a PySpark pipeline for efficient edge weight generation.

Implemented Fast-RCNN to extract nutrition panel with a mAP of 85 from product images.

Utilized Azure OCR to extract nutrition components and their corresponding values with 93% accuracy.

University of Edinburgh, Edinburgh, UK

M.Sc. Data Science — Distinction

Returning to academia to refresh skills and deepen expertise after four years in data science; Edinburgh provided an ideal environment to engage with emerging technologies.

Chose courses like NLP, Extreme Computing, and Financial Data Pattern Recognition, building a foundation to address industry challenges.

Dissertation: built a QA framework for financial analysis, scraping 70GB of annual reports and improving results by 16% over RAG methods.

University of Delhi, Delhi, India

B.Tech IT & Mathematics — 89.5 / 100

Completed B.Tech in IT and Mathematics with a minor in Computational Biology, gaining a strong foundation in math, programming, and practical applications through project-based learning.

Internships including one at DRDO refined skills; courses in NLP and AI sparked a career interest in data science.

Active in the robotics society, managed social media, and organized a data science hackathon — started full-time career at Spoonshot after a final-semester internship there.