Léa Lacoste

• ● Sole owner of the full AI stack for an insurance Q&A assistant: document ingestion, PDF extraction (including complex table parsing), vector search, and LLM-based conversational reasoning

• ● Designed and implemented end-to-end RAG pipeline exposed via a FastAPI backend, handling the particular challenges of insurance document structure

• ● Built production LLM chat agents with retrieval backed by OpenSearch, where conversation history was used to build user profiles and personalize search results

• ● Ran large-scale document extraction pipelines on Databricks using locally-hosted models, processing hundreds of thousands of documents

• ● Optimized search relevance and developed scalable data pipelines for automated web scraping and extraction

• ● Stabilized an overloaded ElasticSearch cluster that was regularly hitting resource limits during peak hours, reducing compute costs by ~90% and memory usage by ~75%

• ● Led architectural improvements across search, storage, and event handling to increase performance and reduce AWS spend

• ● Helped launch Tactic Generate, an early RAG product enabling users to upload documents, generate summaries, and query content via natural language