Architecture Blueprint
AI-Powered SNF Referral
Management Platform
End-to-end architecture — from faxed PDFs to admit/deny recommendations
📠
Faxed Referrals
🏥
Hospital Portals (Epic, AllScripts, NaviHealth)
📄
Direct PDF Uploads
📧
E-Fax (Twilio)
🔗
HL7v2 / FHIR Feeds
1
📥 Document Ingestion & OCR
300+ page referral packets → clean structured text · <3 min per packet
📝 Marker
PDF → Markdown with 96% accuracy. Tables, multi-column, headers preserved. LLM-enhanced mode.
PRIMARY OCR · GPL-3.0
🔍 Surya
Transformer-based OCR. 97% similarity vs Google Cloud Vision. 90+ languages, 0.62s/page.
SCANNED DOCS · GPL-3.0
✍️ TrOCR
Microsoft's handwriting recognition. 94.6% accuracy on medical handwritten forms.
HANDWRITING · MIT
👁️ Qwen2.5-VL
Vision-language model. Reads complex pages visually — tables, charts, mixed content. 7B params.
COMPLEX DOCS · Apache 2.0
📐 Docling
IBM layout analysis. Document structure understanding, reading order, section detection.
LAYOUT · MIT
🔧 OpenCV
Preprocessing: deskew, denoise, contrast enhance, upscale to 300 DPI. +5-10% accuracy on faxes.
PREPROCESSING · Apache 2.0
2
🧬 Clinical NLP & Entity Extraction
Raw text → structured medical entities (diagnoses, meds, insurance, demographics)
🏥 scispaCy
Allen AI clinical NLP. Medical NER: diagnoses, medications, procedures, anatomy. Battle-tested.
CORE NLP · Apache 2.0
🧠 BioBERT
Fine-tuned on n2c2/i2b2 datasets. ICD-10 code extraction, clinical relation detection.
NER MODEL · Apache 2.0
💊 MedXN
Mayo Clinic medication extractor. Drug names → RxNorm normalized codes. Interaction checking.
MEDICATIONS · Open Source
🔒 Presidio
Microsoft PHI de-identification. 94% recall on patient names. HIPAA compliance layer.
DE-ID · MIT
🏗️ FHIR Resources
Map entities to FHIR Patient, Condition, MedicationStatement, Coverage. Interoperable output.
STANDARDS · Open Source
3
🧠 Intelligence Engine — Fine-Tuned LLM + RAG
Clinical reasoning, facility matching, financial analysis · Custom-trained on healthcare data
⚡ Qwen 2.5 (32B)
Base model fine-tuned with QLoRA on MIMIC-III/IV clinical data. Top clinical reasoning scores.
BASE MODEL · Apache 2.0
🔧 LLaMA-Factory
Fine-tuning framework. QLoRA r=64, 72% MedQA accuracy. Trains 32B model on single A100 in ~12hrs.
FINE-TUNING · Apache 2.0
📚 LlamaIndex + LangChain
Hybrid RAG. Facility criteria, payer rules, drug DBs, ICD-10 codes injected per query.
RAG FRAMEWORK · MIT/Apache
🗄️ Weaviate
Vector database. Multi-tenant (per facility), hybrid search, on-prem HIPAA deployment.
VECTOR DB · BSD-3
📋 SGLang / Outlines
Guaranteed valid JSON extraction. FHIR-compatible schemas with per-field confidence scores.
STRUCTURED OUTPUT · Apache 2.0
📊 RAGAS + DeepEval
Evaluate faithfulness, relevancy, hallucination rate. Custom clinical accuracy metrics.
EVALUATION · Apache 2.0
4
🤖 Multi-Agent Decision System (LangGraph)
Specialized agents collaborate to produce admit/consider/deny with transparent reasoning
🔀
Triage Agent
Classify urgency & route
→
🩺
Clinical Agent
Risk assessment & care needs
→
💰
Financial Agent
PDPM, insurance, med costs
→
✅
Criteria Agent
Facility match via RAG
→
📝
Explanation Agent
Reasoning + page citations
5
📤 Decision Output & Integrations
Structured recommendations pushed to EHRs, dashboards, and clinical review queues
⚖️ Admit / Consider / Deny
Transparent recommendation with confidence score, reasoning chain, and source page citations from the referral packet.
PRIMARY OUTPUT
📊 Clinical Summary
Single-page patient overview: diagnoses, medications, risks, care needs, financial projections.
SUMMARY
🔌 EHR Push
PointClickCare + MatrixCare integration. FHIR R4 resources. Bidirectional sync.
INTEGRATION
👨⚕️ Human Review Queue
Low-confidence items routed to clinicians. Override feedback loops back into model training.
HUMAN-IN-THE-LOOP
⚙️
🏭 Production Infrastructure
HIPAA-compliant, scalable, monitored — 240+ packets/day on a single GPU
🚀 vLLM
Model serving. 24x throughput vs alternatives. AWQ quantization. OpenAI-compatible API.
SERVING · Apache 2.0
☁️ HealthStack
Open-source IaC for AWS. HIPAA Terraform modules: VPC, encryption, audit logging, BAA-ready.
INFRASTRUCTURE · OSS
📈 MLflow + W&B
Experiment tracking, model registry, A/B testing. W&B has HIPAA BAA for enterprise.
MLOPS · Apache/Comm
🖥️ NVIDIA L4 GPU
24GB VRAM, $1.50/hr on AWS. Single GPU handles full pipeline. Scale to multi-GPU as needed.
COMPUTE · $800/mo
🌐 FastAPI
Async API layer. OAuth2/OIDC auth. Webhook callbacks. Rate limiting via Celery + Redis.
API · MIT
📊 Grafana + Prometheus
Observability: latency, accuracy drift, error rates, GPU utilization, model performance.
MONITORING · OSS
18
Weeks to Build
96%
OCR Accuracy
<3m
Per Packet
240+
Packets/Day
$2.4K
Monthly Cost
100%
Open Source Core