Danielle Timmins

"Data for Good" only works if the people we're trying to serve can trust what they're being served - and if the system is designed for the messiness of real life: uncertainty, sensitive topics, unequal access, and wildly different levels of digital literacy. In this session, I'll share how we approached building Kozani, a practical, lightweight AI experience designed around trust-first information rather than "best-guess" generation. Instead of treating an LLM like an oracle, we treated it like a coordinator - one that must (1) clearly separate what it knows from what it retrieves, (2) route users to vetted sources when stakes are high, and (3) respond with dignity when the "right" answer is actually support, not certainty. We'll walk through the blueprint: a simple chat interface, a deliberate "search trusted sources" path for grounding, and a model-choice layer that balances cost, latency, and reliability depending on the use case. From there, we'll go into the part teams often skip: how you prove it's helping. I'll cover practical evaluation methods (harm-aware QA, source-quality checks, escalation patterns, and human feedback loops) and what changes when your users are not "power users" - but everyday people looking for guidance, services, or clarity.