AI Agents SF #14 — Healthcare Agents
The AI Agent Meetup is brought to you by Neo4j, The AI Alliance and AI By the Bay.
Agenda:
5:00 pm Doors open. Registration & Networking
5:30 pm Welcome & Introductions
5:45 pm Talks
(1) From Guardrails to Guardians: Architecting Autonomous Oversight for Agentic Healthcare AI by David Talby, John Snow Labs
(2) "The Patients Voice AI Left Behind — Why We Spent Four Months in Clinics and Rebuilt Everything" by David Sarabia, ClinicaMind
(3) Closing the Care Gap: Building Agentic AI for Environments Where Data Can't Move, Siam Tonmoy, Actian
(4) When AI Can't Be Wrong: Running Healthcare Agents for Hundreds of Thousands of Patients a Month, Gediminas Pažėra, Develop Health
7:15 pm Q&A
7:30 pm Open Mic - 30 seconds each
7:45 pm Networking w/Pizza and Sodas
8:30 pm End
Talk Details and Speaker Bios:
(1) From Guardrails to Guardians: Architecting Autonomous Oversight for Agentic Healthcare AI by David Talby, CTO at John Snow Labs
As LLM-based agents move from simple chat interfaces to autonomous, tool-enabled workflows in high-stakes domains like healthcare, traditional safety methods are proving insufficient. Static guardrails cannot account for the non-deterministic tool use and long horizon reasoning required for real-world tasks. This talk explores the architectural shift from passive guardrails to active guardians, an oversight design pattern for continuous monitoring, safety, and regulatory compliance. We’ll cover:
· The Guardian Agent (Runtime Oversight): Architecting real-time oversight loops that perform continuous red teaming while agents are live in production. Guardians monitor for emerging bias and safety issues, performance drift, reasoning failures, and regulatory compliance.
· Healthcare-Specific Testing (Gatekeeping): Building CI/CD release gates to automate the evaluation and detection of flaws, biases, and unsafe behaviors during development. We will dive into the necessity of healthcare-specific test suites for challenges like clinical cognitive biases, regulatory hardening, and explainability for patients and clinicians.
· Open-Source Foundations (Benchmarking): Utilizing MedHelm and HealthAdminBench (Stanford) to move beyond medical licensing exams and evaluate agents via real-world, non-linear tasks including complex clinical, financial, and administrative workflows.
Through real-world Pacific AI case studies in remote patient care and science communication, we’ll share best practices and methodologies developers can use to move agentic systems from experimental prototypes to production-ready, governed healthcare assistants.
David Talby is the CEO at John Snow Labs and Pacific AI, helping companies apply artificial intelligence to solve real-world problems in healthcare and life science. David has extensive experience building and running web-scale software platforms and teams – in startups, open-source projects, for Microsoft’s Bing in the US and Europe, and to scale Amazon’s financial systems in Seattle and the UK. David holds a Ph.D. in Computer Science and Master’s degrees in both Computer Science and Business Administration. He was named USA CTO of the Year by the Global 100 Awards in 2022, Game Changers Awards in 2023, and ACQ5 Global Awards in 2025.
(2) "The Patients Voice AI Left Behind — Why We Spent Four Months in Clinics and Rebuilt Everything" by David Sarabia of ClinicaMind
The best healthcare AI isn't built in a polished demo and a pitch deck — it's built doing rounds at the hospital, sitting with clinical staff between patients, and watching workflows break in real time. We moved into four clinics for four months — learned why providers wanted nothing to do with another surface-level AI. It taught us that everything we were building was wrong. Using design thinking with clinical teams — building with them, not for them — we rebuilt from scratch. What emerged was ClinicaMind OS, a collaboration-first platform that flips how medical systems are built. The most resistant staff became our biggest champions. And our voice-first approach revealed the hardest problems worth solving — making AI work for a 78-year-old Spanish-speaking grandmother with early-stage dementia. This is the unfiltered story of ClinicaMind: from idea, to hackathon prototype, to early failure, to living in hospitals to success in production. Why the future of healthcare AI is collaborative, not autonomous. And why, in a market flooded with AI agents, clinical research — not features — is the only moat that matters.
David Sarabia, CEO and co-founder of ClinicaMind, the company building Patient Memory as healthcare's missing layer. A serial health tech entrepreneur with over a decade in the industry and two prior exits, including a unicorn, David's path into healthcare wasn't a career move. It was personal. After years of success in tech including a unicorn exit to Insight Partners, addiction and homelessness stripped everything away. Recovery led him to build inRecovery, a data-driven addiction care platform that served health systems like Northwell and UHS. Then his father, an immigrant now in his 90s, began losing his memory to early dementia, and David watched medicine lose its memory of him at the same time. Every appointment started from zero. ClinicaMind exists to close that gap. He is joined by co-founders Dasha Sarabia (COO) and Dr. Nilay Shah (CMO), a practicing neurologist and former Chairman of the American Academy of Neurology. The company's research is developed with AI Alliance in partnership with IBM and Meta, David has guest lectured on applied AI in healthcare at MIT, Caltech, and Harvard.
(3) Closing the Care Gap: Building Agentic AI for Environments Where Data Can't Move, Siam Tonmoy, Actian
Building AI agents for clinical workflows means working inside hard constraints. Patient data can't leave the building, retrieval errors have real consequences, and most agentic stacks weren't designed with either of those things in mind. In this talk, Siam walks through the Care Transition Copilot: an agentic system built to assemble patient context, detect risk signals, and generate actionable insights for care teams supporting patients at home. The focus is on what those constraints actually force you to solve: how you architect a RAG pipeline when data sovereignty is non-negotiable, how you keep retrieval accurate enough for clinical decision support, and what it takes to move from a working prototype to something a care team depends on. Relevant for anyone building agents in regulated or data-sensitive environments.
Siam Tonmoy is a Developer Advocate at Actian, focused on agentic systems and RAG pipelines. He builds at the intersection of AI infrastructure and applied development, with hands-on experience deploying retrieval-augmented systems in data-sensitive environments.
On the Vector DB sensitivity question: in most domains it isn't that sensitive, but in healthcare it is, because the data can't move. Cloud-native vector databases assume you can bring data to the index. In clinical environments that's often a compliance non-starter. The talk uses Actian VectorAI DB because local and on-prem deployment was a hard requirement for this use case, not a preference. The architectural decisions that follow from that constraint are what the talk is really about.
We are looking for more talks on the fundamental questions of Agentic AI platforms, framework and protocols, especially whether a platform layer will hold. Comment with proposals or email us!
(4) When AI Can't Be Wrong: Running Healthcare Agents for Hundreds of Thousands of Patients a Month, Gediminas Pažėra, Develop Health
At Develop Health I build the AI systems that automate insurance prior authorization, running in production for over 100,000 patients a month. This talk is a case study of that system and the lessons we picked up building it. I'll touch on many aspects of running AI in production: evaluating properly (and the eval bugs that quietly inflate your accuracy numbers), making good use of human annotations, cutting costs with cheaper models without giving up quality, being careful with patient data, and some of the features we're most proud of, like teaching the system to catch its own mistakes before they reach a payer. Whether you're building in healthcare or anywhere else mistakes are expensive, you'll leave with practical tips you can use.
Gediminas Pažėra is an AI Engineer at Develop Health, where he builds the generative AI systems behind automated prior authorization: clinical record understanding, payer form answering, denial risk detection, and AI-drafted appeals, all with human review before anything reaches a payer. Previously founding ML engineer at Lighten AI and co-founder of a health tech company. DPhil in Chemical Physics, University of Oxford.