Agentic AI is everywhere right now — writing code, answering questions, routing tasks, and stitching together workflows like an eager new intern. But in healthcare, an overconfident agent that “just makes something up” isn’t being creative. It’s being dangerous.

Over the past year, I’ve built a set of healthcare agents and decision-support tools — NephroCompass, Carelytics (Hospital BI), and a Blood Donation DSS — all of which interact with sensitive data and real patient journeys. When you’re dealing with someone’s health, the room for error shrinks fast. Every hallucination isn’t just a bug; it’s a breach of trust.

This session is the honest, behind-the-scenes look at what it actually takes to make agentic AI behave like a responsible teammate instead of a chaotic one. It’s the messy middle — designing systems that stay grounded, reason in context, and hold up under pressure, even when the data is imperfect, emotional, or fragmented.

We’ll talk about:

• Why plain RAG falls short in healthcare, and what actually makes retrieval reliable
• How domain grounding, layered retrieval, and context shaping reduce hallucinations
• Architectures that keep agents safe without turning them into slow bureaucrats
• How to design step-by-step reasoning patterns that clinicians can follow and trust
• Real moments from NephroCompass and Carelytics — where the agent almost went off-track, and how we pulled it back
• How to test, validate, and monitor an agent when your “users” are patients, pharmacists, and care teams who can’t afford mistakes
• Why sometimes the right fix isn’t clever prompting — it’s solid, boring, beautiful data engineering

This isn’t a talk about futuristic magic. It’s about building AI that behaves well in the real world, especially where the stakes are personal and human.

Building agentic AI is fun.
Building agentic AI that people can trust — that’s the real craft.