Unnat Agarwal

Improves insurance content search efficiency by 4x by using inverted indexing by developing an AI autotagger. Achieves up to 90% document classification accuracy and extracts over 80% relevant searchable content by optimizing prompt design and inference pipelines with cost-efficient LLMs. Enables cross-insurer product analysis and filtering using an NLP pipeline with HuggingFace NER, high-throughput Bi-Encoder retrieval, and fine-tuned Cross-Encoder re-ranking against a curated global feature taxonomy. Deployed and integrated the Flask-based tagging service leveraging Gemini AI with LifeRay CMS on the internal network, enabling continuous background content ingestion and indexing.

Implemented the backend for Mario, an app supporting event posts, stories, and merchandise sales with a points-based rewards system with 300+ users. Developed Emily, a MERN application for creating interactive chatbot-style forms with exportable Excel reports with 100+ satisfied users.

Shortened time-to-hire by developing the core AI Sourcing Engine pipeline, which stored candidates sourced from platforms like LinkedIn in a Weaviate vector database. Optimized recommendations—delivering up to 2x more interested candidates per recruiter, by developing an advanced reranking system using XGBoost and PyTorch, powered by extensive feature engineering to match candidates to job descriptions with high precision. Cut sourcing costs by over 50% by reducing time spent on unsuitable profiles through online learning (MAML-based) to continuously adapt the reranker to recruiter preferences in real time, increasing matching accuracy over time. Hosted models on AWS SageMaker with model storage in AWS S3 for scalable and reliable inference.