Preventive care campaign
Patient list
2,400MBMaria B.Screening overdueReady
JTJames T.Follow-up dueQueued
ALAna L.Screening overdueQueued
Healthcare AI solution portfolio
Close more preventive care gaps—with outreach that can finish the work.
A production-shaped patient outreach agent that identifies overdue patients, engages them by SMS, answers approved questions, books real slots, and writes the outcome back—while keeping safety-critical decisions under workflow control.
Portfolio case study · synthetic patient data · POC in development
Patient-facing AgentWorkflow + bounded model loop
SMS conversationMaria B. · identity verified Live
Hi Maria, Harbor Health has two screening appointments available next week.
Tue, July 14 · 3:30 PM
Tuesday works for me.
Appointment bookedEHR outcome updated
Business outcomes
Designed around the outcome that matters.
This is not a chatbot success metric. The north star is whether an eligible patient’s preventive-care gap is closed within 60 days.
Grow patient engagement
Make timely outreach feel personal without adding another manual call list for staff.
Book more screenings
Remove portal and phone-tag friction by completing appointment booking in the same SMS conversation.
Release staff capacity
Automate repetitive outreach while routing clinical, administrative, and patient-requested exceptions to people.
Close care gaps
Connect outreach activity to appointment and EHR outcomes instead of stopping at messages sent or replies received.
Business case v1.0
A measurable pilot, not a vague AI promise.
Harbor Health’s simulated discovery findings become a testable 60-day pilot hypothesis. Targets are planning inputs—not claimed results.
Low 20%Base 40%High 50%
Customer-approved labor-reduction scenariosAgentic workflow
One constrained agent. One complete care journey.
A deterministic workflow controls the sequence. The model is used only where language understanding and bounded tool choice add value.
03
Pre-rules → Model → Tool Gate
Understand, answer, and route
Rules handle opt-out and risk first. A bounded model loop then understands patient intent and may use only approved knowledge while the workflow retains control.
What the POC proves
Small enough to demo. Shaped for production.
Is this Maria?Yes, it is.
Privacy-aware SMS outreach
Neutral first contact, identity verification, inbound normalization, deduplication, and deterministic opt-out.
Approved FAQharbor-faq-prep-001 · v1.0Grounded ✓
Approved knowledge answers
The model may rewrite retrieved, versioned FAQ content. No result means no answer from model memory.
Tue 3:30Thu 4:00BOOKED
Real-time mock scheduling
Separate read and write tools ensure displayed slots are real and booking occurs only after explicit confirmation.
“I have chest pain and trouble breathing.”Staff task created
Clinical escalation
Clinical escalation
Clinical safety escalation
High-risk language triggers fixed guidance and human handoff before the booking tool can be exposed.
Solution architecture
Agent = Model + Healthcare Harness
The model understands language and chooses among currently allowed actions. The harness provides the state, context, workflow, tools, safeguards, reliability, and human ownership needed for enterprise use.
Patient Outreach Agent
MODELMicrosoft Foundry
Prompt Agentreason · understand · propose
Prompt Agentreason · understand · propose
+
HARNESSHealthcare Agent Harness
Deterministic workflowContext builderState & memoryPre / post guardsDynamic tool permissionsRetry & timeoutAudit & traceHuman handoff
lookup_approved_knowledgesearch_slotsbook_appointment
Logical components shown here do not imply independently deployed microservices.
Trust & control
The agent has judgment—but never unlimited authority.
Every tool call passes through workflow state and typed validation. Safety-critical actions stay deterministic, observable, and recoverable.
Minimum necessary context
Screening details and scheduling tools remain hidden until the identity step is satisfied.
Dynamic authorization
The model sees only the tools allowed by the current state and still cannot execute them without server-side validation.
Idempotent booking
Unknown results are persisted and escalated rather than retried into a duplicate appointment.
Safe handoff
Clinical risk, complex requests, and explicit staff requests stop automation and create a contextual queue item.
Design evidence
Enterprise rigor, made visible.
Agent trace · maria-b-00421.84s total
WORKFLOW STATEAWAITING_CONFIRMATION
MODEL DECISIONBOOK_APPOINTMENT
TOOL GATEALLOWED
TOOL RESULTCONFIRMED · apt_9281
› book_appointment(patient_id, slot_id, idempotency_key, confirmation_evidence) 200 OK
Planned Azure POC
A credible path from synthetic demo to client-owned pilot.
The POC validates technical feasibility in a demo environment. Real patient use moves to a customer-controlled pilot with compliance, security review, and operational readiness.
01Microsoft Foundry02FastAPI03React04PostgreSQL05Container Apps06Managed Identity07Key Vault08Application Insights
Synthetic data, Mock EHR/Scheduling/SMS, one complete journey and one safety handoff.
Customer Azure, real integrations, treatment/control measurement, compliance controls and SLA.
Scale, resilience, disaster recovery, operations, additional campaigns and channels.
Interactive demo
Watch one journey from care gap to booked appointment.
The final recording will show the golden path, one clinical escalation, and the agent trace that proves the right tools were—and were not—called.
01 · Golden path02 · Safety handoff03 · Trace evidence
Patient Outreach Agent · Demo06:00
DEMO RECORDING RESERVEDVideo will be added after the POC build.Reserved path: /public/demo/patient-outreach-agent.mp4
My role
AI Agent Solution Architect
End-to-end ownership across discovery, value engineering, agent design, Azure architecture, risk controls, evaluation, and the POC roadmap.
Customer discovery & value alignmentAgentic workflow & harness engineeringTool contracts & transaction safetyEvaluation & failure-mode designAzure POC & pilot roadmapROI / TCO framework