AMIT ARORA

Developing and optimizing production ML and NLP solutions for email security, focusing on model performance, inference efficiency, and threat detection. Building agentic AI workflows to automate topic creation through intelligent, example-driven processes, while engineering scalable, cost-efficient ML pipelines for production deployment.

Key Achievements:

• High-Precision Email Topic Classification: Developed and deployed a SetFit-based model to classify emails across 18 topics, achieving over 90% precision with strong recall. Productionised the model using an AWS SageMaker endpoint.
• ⁠120%+ Higher Throughput & Cost-Efficient NER Inference: Evaluated multiple NER models and built an end-to-end inference pipeline from scratch. Optimized context length, converted the selected model to ONNX , and benchmarked instance types to increase throughput by over 120%, reduce latency, and improve serving cost efficiency.
• ⁠Enhanced Detection & 3× Higher Trigger Rate: Retrained and enhanced the NTP model by introducing new features and removing stale features, increasing trigger rates from 8 to 25 per million while targeting false-positive reduction.
• Agentic AI for Automated Topic Creation: Contributed to Topic Foundry, an agentic AI workflow designed to enable users to create new email topics from a small number of examples. Established an end-to-end pipeline for four initial topics as a foundation for automated topic creation.
• Optimized Production Embeddings: Evaluated candidate embedding models across embedding dimensions, token lengths, classifier architectures, predictive performance, and inference latency to identify the best trade-off for production deployment.
• Customer-Driven Detection Improvements: Investigated customer reported false positives and production alerts to diagnose detection issues, support resolution of customer concerns, and improve email threat detection quality.

Experienced Data Scientist with proven success in leading and managing cross-functional teams in the pharmacovigilance domain. Expert in developing and deploying machine learning solutions, particularly in NLP, to streamline business processes and enhance operational efficiency. Adept at collaborating with stakeholders to translate business needs into actionable data-driven solutions, while consistently delivering high-impact results.

Key Achievements:

• Engineered and automated data pipelines using AWS Athena, Glue, and Lambda functions, enabling real-time analytics for a business-critical application.
• Designed and deployed an advanced NLP pipeline for Entity Extraction, Resolution, Assertion, and Classification models, with the overall accuracy of 86%. Integrated with the HALOPV product, this solution reduced users' processing time by 75%, significantly enhancing case management efficiency.
• Implemented QnA and LLM models to extract patient and reporter information from unstructured text, achieving 82% accuracy in data points extraction.
• Containerized NLP pipeline models using Docker, creating scalable, on-demand endpoints that improved accessibility and reduced deployment time.
• Leveraged AWS services (EC2, Sagemaker, Lambda, Step Functions) to deploy and integrate the NLP pipeline, optimizing cloud infrastructure costs and enabling seamless, scalable model operations.
• Developed and deployed a customer-centric chatbot leveraging Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) techniques to provide personalized support.
• Developed interactive reports and dashboards using SQL, enhancing data accessibility and providing real-time insights for stakeholders.

NLP Engine

• Developed script for automated entity annotation in documents, improving training data generation efficiency.
• Built an end-to-end NLP engine integrating a SpaCy NER model and Regex-based entity extraction, using a custom Python algorithm for resolution.
• Deployed the NLP model in AWS Sagemaker via Docker for scalable, cloud-based access.
• Created Naive Bayes model to classify insurance claims as casualty or property, enhancing claim categorization process.

Digital Desk Claim Triage

• Designed Neural Network to predict insurance claim severity, leveraging data analysis on loss description texts to inform decisions.
• Visualized insights using word clouds and t-SNE plots to enhance understanding of claim patterns.
• Developed a time series model to forecast the cumulative number of insured claims, improving claim management over time.