Trust & build · How it fits together
Platform Architecture & Rollout
A modular platform that sits next to the EMR instead of replacing it: standards-based integration, a central data layer, AI services behind guardrails, and a 90-day rollout plan built around measurable outcomes.
modular servicesstandards-firstcloud-agnosticsecure by defaultmeasured rollout
Reference architecture
Every layer is replaceable; standards keep it vendor-neutral
Channels
Practice website & smart intakePatient portalSMS / IVR / voicemailStaff consoleFax & email
Applications
Prior Auth CopilotDocument AIClinical ScribeClaim ScrubberSchedulingReferralsAI AgentCommand Center
AI & services
LLM gateway (Claude / OpenAI, BAA, ZDR)Clinical NLPOCR / Document AIML models (no-show, denial risk)Rules engineWorkflow orchestrator
Integration
FHIR R4 gateway (SMART)HL7 v2 MLLP listenerX12 via clearinghousePayer PA APIs (Da Vinci)Fax / SMS providersWCB Medical Portal
Data
Operational PostgresFHIR storeWarehouse (star schema)Object storage (raw)Feature storeImmutable audit log
Cross-cutting: SSO/MFA · RBAC · encryption · audit logging · PHI redaction · monitoring & alerting · BAAs
99.95%
uptime target (SLO)
multi-AZ, health checks, auto-restart
< 200 ms
p95 API latency
caching + read replicas for dashboards
0 lost messages
HL7 & X12 delivery
durable queues, retries, dead-letter + replay
< 1 h / 15 min
RTO / RPO
PITR backups, infra as code, DR runbook
Languages
- • TypeScript (Next.js, Node)
- • Python (FastAPI, pandas, scikit-learn, PyTorch)
- • Go (high-throughput HL7 listener)
- • Java (Mirth Connect channels) · C# (.NET vendor SDKs)
- • SQL
AI / ML
- • Claude & OpenAI APIs
- • LangChain-style tool orchestration
- • spaCy / transformer NER
- • TensorFlow / PyTorch for custom models
- • Evals & guardrails
Healthcare standards
- • HL7 v2.5.1
- • FHIR R4 · US Core · SMART
- • X12 5010 (270/271, 276/277, 278, 837P, 835)
- • ICD-10-CM · CPT · LOINC
Cloud & infra
- • AWS (HIPAA-eligible services) or Azure / GCP
- • Docker · Terraform · GitHub Actions
- • Queues + dead-letter + retries
- • OpenTelemetry monitoring
First 90 days
Start with the people doing the work, then automate what measurably helps
Days 1–30
Listen, measure, quick wins
- Shadow front desk, auth, billing and providers at each site
- 2-week time study → baseline for every KPI
- Read-only EMR / clearinghouse access and data map
- Ship: pre-visit checklist SMS and fax-intake classification
Days 31–60
Revenue cycle automation
- Prior Auth Copilot for the top 5 CPTs by volume
- Claim scrubber rules from 12 months of 835 denials
- Central pipeline + Command Center dashboard
- Weekly KPI review with operations
Days 61–90
Scale & intelligence
- No-show model live with waitlist backfill
- Clinical scribe pilot with 1–2 providers
- AI agent for rescheduling & status questions (EN/ES)
- 90-day results vs. baseline → next-quarter roadmap
From the codebase
Real excerpts from this demo
// HL7 v2 → FHIR (lib/hl7.ts, simplified)
export function parseHl7(input: string): Hl7Message {
const lines = input.replace(/\r\n|\n/g, "\r").split("\r").filter(Boolean);
const fieldSep = lines[0][3]; // MSH-1
const [comp, rep, esc, sub] = lines[0].slice(4, 8); // MSH-2
const segments = lines.map((line) => {
const [id, ...fields] = line.split(fieldSep);
return { id, fields: fields.map((raw) =>
raw.split(rep).map((r) => r.split(comp).map(unescape))) };
});
return { segments, type: field(segments[0], 9), controlId: field(segments[0], 10) };
}