Looking for a job

Shobhit Maheshwari

@shobhitmaheshwari

Message

I build production AI systems for insurance automation, document intelligence, and multimodal extraction.

India

Message

At Roadzen, I build production AI systems for motor insurance claims, including Autospect, a Claude-powered agent that automated 80% of pre-inspection workflow and cut claim intake from hours to minutes. I also developed computer vision and multimodal extraction pipelines reaching 93% AP50, 97.8% vehicle identity accuracy, and up to 90% field extraction accuracy.

I've built end-to-end ML infrastructure with FastAPI, Airflow, Jenkins, TorchServe, and Docker, while improving insurance QA to 95% accuracy and reducing claims processing from 40 minutes to under two. My work spans computer vision, RAG, LLM fine-tuning, and financial QA research at the University of Edinburgh.

Experience

Work history, roles, and key accomplishments

Current

Lead Data Scientist

Current

Roadzen

Apr 2025 - Present (1 year 5 months)

Built Autospect, a Claude-powered motor insurance claims agent that automated 80% of the pre-inspection workflow, reducing claim intake turnaround time from hours to minutes. Engineered a from-scratch instance segmentation pipeline using a frozen DINOv2 backbone and Mask2Former head for car-part segmentation, achieving ~93% AP50 on the test dataset.

Claude DinoV2 Mask2Former Instance segmentation Qwen2.5 VL LoRAs Quantization Gemini Computer Vision NLP

Research Assistant

University of Edinburgh

Aug 2024 - Jan 2025 (5 months)

Scraped and processed 70GB of historical FTSE All-Share annual reports for financial QA research. Benchmarked GPT against 4-bit quantized Llama and Mistral models on QA tasks, with GPT outperforming both by ~12% in answer accuracy.

Data Scraping Data Processing GPT LLaMA Mistral RAG MapReduce Quality Systems

Senior Data Scientist

Roadzen

Jun 2020 - Aug 2023 (3 years 2 months)

Built end-to-end ML pipelines for data ingestion, training, and model serving using FastAPI, Airflow, and Jenkins, enabling automated retraining and deployment of in-house ML models. Enhanced an insurance policy QA system using GPT-based retrieval, achieving 95% answer accuracy.

fastAPI Airflow Jenkins GPT Mask RCNNs PyTorch TorchServe ML Pipelines Model Serving

Data Scientist

Spoonshot

Jan 2019 - Jun 2020 (1 year 5 months)

Designed a weighted DeepWalk-based graph model using ingredient co-occurrence and flavour pairing theory to generate novel flavour combinations. Scaled graph edge weight generation to over 100 million ingredient combinations using PySpark.

DeepWalk Graph Models PySpark Fast R CNN Computer Vision Flavor Pairing Education