Student Insurance Navigator

Many US college students have to buy their own health insurance. PhD students who turn 26 and lose a parent's plan are a common example. Most of them choose by guessing. Our product is an AI agent that helps them pick a plan that fits their real needs. The agent asks a few short questions at a time instead of making students fill out one long form. It asks about doctor visits, prescriptions, preferred doctors, and what matters most: low premiums, a low deductible, or a large network. It then compares plans using two sources: benefit details from official plan documents, and real prices insurers have negotiated with providers. If a student asks about an insurer we don't have, the agent searches for it online. The agent is composed of an LLM(supported by AMD) + a Python calculator to make precise insurance calculations. We use Harness to support the LLM's ability and database as a support for the Python calculator. Our harness keeps the open-source model on a short leash. A main prompt defines the agent as an insurance advisor and sets ground rules, each one fixing a mistake we saw in testing. Skills package instructions with tools. For example, the questionnaire skill holds a fixed question bank, and the model can only pick from it, so it can't make up questions. The tools let the model look up a plan rule, run the cost calculator, explain an insurance term, or search for an insurer we don't have. Hooks are Python checks that run around every model step. Because health-related agents are very sensitive to data accuracy and privacy, the hooks can both secure and validate data. Our agent relies on two databases. Qwen 3.8-27b model on the AMD API reads the benefit tables in each plan PDF and turns every row into a structured rule. Each rule covers one service on one network and notes whether prior authorization is needed. Each rule also records the page it came from and is verified by us. If a rule is unclear or hasn't been reviewed, the calculator won't use it.