Job Description & Details
This is a heavy-hitting 12+ month contract for someone who speaks both fluent P&C insurance underwriting and advanced LLM prompt engineering. You are going to be bridging the gap between legacy insurance workflows and cutting-edge generative AI models. If you don't have deep, foundational US-based insurance experience mixed with hands-on AI model interaction, this one will be a stretch.
What You'll Actually Be Doing
Your day-to-day will revolve around building, testing, and refining custom prompts, managing tokenization constraints, and handling content chunking for massive underwriting datasets. You will sit directly between the IT teams and the underwriting stakeholders, taking complex submission and evaluation processes and translating them into automated, AI-driven ingestion pipelines. Expect to spend a lot of time debugging model outputs, ensuring risk assessment logic is tight, and integrating LLMs smoothly into existing enterprise architectures.
The Core Tech Stack
You need absolute mastery over prompt engineering and design, specifically with tools like Claude, GitHub Copilot, and the broader copilot suite. Familiarity with open-source, open-weighted, and closed LLM architectures is a must. On the data side, you'll need working knowledge of vector databases, graph databases, and ontologies, alongside solid programming fundamentals in Python, R, or Java to handle the heavy data manipulation required for content chunking.
Interview Expectations
Expect the hiring team to hand you a messy, unstructured insurance submission document and ask you to live-engineer a prompt that extracts specific risk factors without hallucinating. They want to see how you handle token limits and context windows on the fly. You'll also likely be grilled on your strategy for chunking dense policy documents and how you query vector versus graph databases for ontology mapping, so have real-world architectural examples ready to draw from.
Application Advice
Make sure your resume leans heavily into your 15+ years of US-based experience, specifically highlighting where you blended P&C insurance domains with GenAI implementations. Don't just list Python and Java; explicitly call out keywords from the JD like content chunking, tokenization, vector databases, and Claude/Copilot integrations so you easily clear their ATS keyword filters.