MamaGuard

• 

  MamaGuard

• 

  MamaGuard Agents

Inspiration

The US has the worst maternal mortality rate in the developed world, and ~80% of those deaths are preventable. The gap is widest for Black, rural, and Spanish-speaking patients on Medicaid — exactly the population where compound risk and coverage fragility stack. Postpartum is the most dangerous window — and the one where care continuity falls apart fastest. The signal is almost never one lab or one missed visit; it's the combination of an elevated BP, a borderline HbA1c, a lapsed Medicaid coverage, a language barrier, and a clinician decision made two weeks ago that nobody re-read. Rules engines flag each piece in isolation. Clinicians don't have 20 minutes per chart to fuse them. I built MamaGuard for the gap between those two facts.

What it does

Meet Maria: postpartum, Stage 2 hypertension, diabetes-range HbA1c, Spanish-only, Medicaid coverage lapsed last week. MamaGuard is a FHIR-native maternal-child care agent that gives her clinician continuity, synthesis, and approval-gated action in one workflow:

• Recalls prior clinician decisions from FHIR DocumentReference so a fresh session knows metformin was already declined for GI intolerance — and doesn't suggest it again.
• Synthesizes compound risk across BP trend, HbA1c, pregnancy history, housing, language, and Coverage gaps into a single urgent narrative: "Stage 2 postpartum HTN + diabetes-range HbA1c + Medicaid gap + Spanish-only support — call her today."
• Writes back a pending Goal + CarePlan + CommunicationRequest — and waits for the clinician to approve in Prompt Opinion before anything is committed.
• Hands off mother → newborn by following the linked patient resource into pediatric follow-up without a manual chart switch.

How I built it

• Agent runtime: Google ADK, deployed to Azure Container Apps (ca-mamaguard + ca-hapi-fhir HAPI R4) for the live demo.
• MCP server: exposes FHIR read + write-back tools so any agent — not just MamaGuard — can use the continuity layer.
• A2A v1 wire format aligned with Prompt Opinion's SendA2AMessage, so MamaGuard is invokable as a BYO agent from PO chat.
• Plan-mode approval gate: the agent emits a structured pending bundle; PO renders it; clinician approval flips the writes to active state.
• Model backends: production runs gpt-5.4; benchmarks swap in Nemotron-3-Super-120B and Gemma 4 models hosted locally on a DGX Spark, with DeepSeek v3.2 as LLM-as-judge.
• SHARP FHIR extension carries the patient ID and FHIR token through the MCP and A2A boundaries, so context survives every hop.

Challenges I ran into

• Saying no to scope. Two weeks out I pivoted from benchmark-maximalism to patient-action storytelling. Cutting good work hurt.
• Memory window vs. demo entropy. Every agent run writes a DocumentReference, which pushes Dr. Kim's prior memory note out of the fetch window — so I built a pre-record cleanup ritual.
• Plan-mode in a chat surface. Getting a pending CarePlan bundle to render and round-trip approval through PO's BYO chat took several iterations of the orchestrator prompt.
• Infrastructure realities. A2A v1 (not v0.3) required dropping deprecated agent-card fields and re-declaring SMART scopes; live deployment landed on Azure Container Apps after a late Cloud Run pivot, with HAPI on in-memory H2 that requires a non-empty fhirToken.

Accomplishments that I'm proud of

• A clinician-approved write-back loop that actually works end-to-end in Prompt Opinion's chat surface, not a screenshot.
• 100% clinician-review flagging on URGENT/HIGH cases in safety verification — nothing dangerous gets auto-committed.
• 93.1% Tier-2b end-to-end pass rate (45/47) on the April 17, 2026 submission run — against a live HAPI FHIR server, not mocks.
• +30% average lift over a naive LLM on the AI-factor benchmark, judged by DeepSeek v3.2 across clinical accuracy, risk, safety, completeness, and output quality.
• A dual-track submission: a usable MCP server other agents can pick up, plus a fully configured A2A agent on the PO marketplace.

What I learned

• DocumentReference is underrated as agent memory. Continuity that lives inside the FHIR boundary survives session resets, model swaps, and clinician handoffs in a way that vector stores don't.
• The judging unlock is the approval gate, not the model. Every clinician I showed this to relaxed the moment they saw that nothing writes until they say so. That single architectural choice answered most of the safety, feasibility, and liability questions before they were asked.
• Compound reasoning is where rules stop and GenAI starts. Five flat flags is a worklist. One synthesized "Stage 2 postpartum HTN + diabetes-range HbA1c + Medicaid gap + Spanish-only support — call her today" is a clinical decision.
• Narrative beats more tools. I had a stretch goal to add three more FHIR tools the week before submission and chose to write the clinician story instead. It was the right call.

What's next for MamaGuard

• Pilot with a community health clinic focused on perinatal Medicaid populations — the cohort where compound maternal + SDOH risk hits hardest.
• SMART-on-FHIR launch into Epic and Cerner sandboxes, replacing HAPI with a real EHR session and SHARP-bridged credentials.
• SFT a smaller model — Gemma-4-26B-A4B on DGX Spark, trained on frontier gold responses — to bring inference cost down to clinic-budget territory and add an on-prem privacy posture for clinics that can't send PHI off-site.
• Extend the pattern beyond maternal-child: pediatric chronic disease, geriatric polypharmacy, and oncology survivorship all have the same compound-continuity shape.
• Open-source the MCP server so any agent author can plug into the same approval-gated FHIR continuity layer I used.

Built With

• a2a
• azure-container-apps
• deepseek
• fhir
• gemini
• gemma
• google-adk
• mcp
• nemotron
• openrouter
• prompt-opinion
• python
• vllm