I need to design the workflow for an LLM orchestrator, integrating with various systems. It should call the customer master DB and policy admin to identify existing clients or handle new quotes. For new cases, I'll clearly differentiate based on line of business (P&C personal auto, commercial, travel, life) and align with CSIO standards. I'll focus on setting up the necessary infrastructure, models, vector DB, middleware, and PoC milestones. Citations will be needed for referencing CSIO API standards. I’ll ensure everything fits within the architecture, including tool-calling patterns and structured forms. Here’s a concrete, LOB aware architecture you can actually brief to a dev partner or internal team. I want to embed an end-to-end AI layer into my insurance website that can greet visitors, guide them through quoting, and quietly run risk-assessment logic in the background. The same engine should draw on three data streams—customer applications we receive in real time, our historical claim records, and wider market-trend datasets—to keep its predictions accurate and current. Here is what the solution must actually do once deployed: • Predict individual risk levels instantly after a prospect submits an application, using supervised models trained on our claims history. • Recommend the most suitable policy and limits automatically, showing the visitor a clear quote while passing the selected product and premium back to our back-office system. • Flag high-risk clients to underwriters with an explanation score so they can override or request extra documentation. • Present all of this via a REST/JSON API that my current React front end can call, and expose a lightweight Python-based admin dashboard where underwriting managers can adjust thresholds and retrain models on fresh data. I work mainly with a Django stack, so a Python solution (TensorFlow, PyTorch, or scikit-learn—whichever you feel is best) will plug in cleanly. The customer-facing chatbot or form assistant can be built with a service such as Rasa or a fine-tuned OpenAI model as long as it routes seamlessly into the same risk engine. Acceptance criteria 1. A reproducible training pipeline with clear data-schema docs. 2. Dockerised API that runs locally and on AWS Elastic Beanstalk. 3. Demo on my staging domain showing a full quote flow from form submission to policy offer, including an underwriter alert when a test “high-risk” profile is entered. 4. Short hand-over video and written setup guide. 5. If this scope is comfortable for you and you can point me to at least one previous project where you automated risk scoring or complex decisioning, let’s talk timings and milestones.