Agents Assemble Healthcare AI Hackathon 2026

ClinicalMem

6-Layer Clinical Safety Pipeline

Persistent, Auditable, Contradiction-Safe Clinical Memory for Healthcare AI Agents

90/90 Tests MCP Protocol A2A Protocol FHIR R4 6 Safety Layers SHA-256 Audit
11
MCP Tools
5
A2A Skills
6
Safety Layers
4
Conflicts Found
Scroll to explore

Live Patient Scenario

SM

Sarah Mitchell

MRN: SM-2026-0847 • DOB: 1959-03-14

2 CRITICAL 2 HIGH

Demographics

Age67
SexFemale
Providers4 Active
Medications7 Active

Active Conditions

Type 2 Diabetes Mellitus
Hypertension
CKD Stage 3
Atrial Fibrillation

Care Team

Dr. Chen Endocrinology
Dr. Patel Cardiology
Dr. Williams Nephrology
Dr. Rodriguez ER Physician
Known Allergy: Penicillin (Documented 2019 — Anaphylaxis)

Safety Findings

All findings detected automatically from FHIR R4 patient data

FINDING #1

Warfarin + Ibuprofen

CRITICAL

NSAID bleeding risk. ER physician prescribed ibuprofen without checking existing anticoagulant therapy. Concurrent use increases risk of GI hemorrhage by 3-6x.

Drug Interaction Pipeline Layer 1 • Deterministic <1ms
FINDING #2

Penicillin Allergy + Amoxicillin

CRITICAL

Allergy cross-reaction. Urgent Care prescribed amoxicillin (beta-lactam) to a patient with documented penicillin anaphylaxis. Cross-reactivity rate: ~2%, but history of anaphylaxis makes this contraindicated.

Allergy Cross-Reaction Layer 1 • Deterministic <1ms
FINDING #3

Declining GFR + Metformin

HIGH

Lab-medication contraindication. eGFR trend: 45 → 38 → 32 mL/min. FDA recommends discontinuing metformin at eGFR <30. Current trajectory reaches threshold in ~3 months.

45
38
32
~26?
STOP LINE
30
Lab-Medication Monitor Layers 1+3 • Deterministic + RxNorm
FINDING #4

Conflicting BP Targets

HIGH

Provider contradiction. Two specialists set incompatible blood pressure targets for the same patient:

Dr. Patel (Cardiology)
<130/80
CONFLICT
Dr. Williams (Nephrology)
<140/90
Provider Contradiction Layer 4+5 • Multi-LLM

6-Layer Safety Pipeline

Each clinical decision flows through all 6 layers sequentially. Green never hallucinates. Purple uses LLMs. Red is the safety gate.

1

Deterministic Table

RULE-BASED

Hard-coded drug interaction pairs and allergy cross-reaction rules. Zero hallucination risk.

2

OpenEvidence API

EVIDENCE API

Mayo Clinic and Elsevier ClinicalKey AI. Peer-reviewed evidence retrieval.

3

NIH RxNorm API

FEDERAL STANDARD

Federal gold standard used by Epic and Cerner. Drug class lookups and interaction validation.

4

Multi-LLM Cascade

AI-POWERED

GPT-5.4MedGemma 27BGemini 3 Flash — Cascading consensus with majority vote.

5

LLM Synthesis

AI-POWERED

Generates patient-specific clinical narratives with explicit evidence citations. Every claim is grounded.

6

Abstention Gate

SAFETY GATE

"I don't know" when evidence is insufficient. The system refuses to guess about patient safety.

Deterministic (never hallucinates)
Evidence APIs (peer-reviewed)
LLM-Powered (grounded)
Safety Gate (fail-safe)

SHA-256 Audit Trail

Tamper-proof Merkle chain — every clinical decision is cryptographically chained to the previous entry

Each block's SHA-256 hash includes the previous block's hash, creating a tamper-proof chain. Any modification to a historical entry invalidates all subsequent hashes. This provides HIPAA-grade audit compliance without a blockchain.

MCP Tools & A2A Skills

11 MCP tools for Claude + 5 A2A skills for agent-to-agent orchestration

MCP Protocol 11 Tools — FastMCP 2.x + SHARP-on-MCP
store_clinical_observation

Store clinical note for patient

recall_patient_context

Hybrid search over clinical memory

check_medication_conflicts

Drug interaction detection (4-tier)

check_allergy_conflicts

Allergy cross-reaction alerts

get_treatment_dependencies

Condition-medication dependency map

get_clinical_audit_trail

Tamper-proof audit chain

summarize_patient_history

Importance-scored patient overview

detect_belief_drift

Cross-provider contradiction detection

ingest_patient_data

Pull all FHIR data into memory

GenAI
explain_clinical_conflict

LLM-grounded conflict explanation with citations

GenAI
clinical_care_handoff

Evidence-grounded care handoff note

A2A Protocol 5 Skills — Google ADK Agent-to-Agent
medication-safety-review

Drug interactions + allergy conflicts

clinical-context-recall

Hybrid search with confidence gating

contradiction-assessment

Scan for conflicting records

care-transition-summary

Structured handoff summary

GenAI
explain-conflict

LLM-grounded clinical rationale with evidence citations

System Architecture

Shared engine powering both MCP and A2A interfaces

FHIR R4 Patient Data
Conditions, Medications, Labs, Allergies
Medical APIs
NIH RxNorm, OpenEvidence, GPT-5.4, MedGemma

Shared Engine

engine/

Clinical Memory
BM25 + Vector + RRF
MIND Kernels
Clinical Scoring
Drug Interactions
4-Tier Pipeline
LLM Synthesizer
Evidence-Cited
SHA-256 Audit Trail (Merkle Chain)
MCP Server
FastMCP 2.x
11 Tools SHARP
A2A Agent
Google ADK
5 Skills A2A
Audit Trail
Hash Chain
SHA-256 Tamper-proof