6 specialized AI microservices unified by an LLM-powered agentic orchestrator — from differential diagnosis and ED trolley forecasting to treatment pathways, clinical note search, and hospital resource optimization. Built for the Irish Health Service Executive (HSE).
A clinician asks a natural-language question through a single conversational interface, and the system automatically classifies, routes, calls downstream services, synthesizes results, and flags safety-critical decisions.
LangGraph + GPT-4o Orchestrator — Port :8100
Differential Diagnosis via GPT-4o CoT
:8005ED Trolley Forecasting via LSTM
:8001Treatment Pathways via BERT
:8002Clinical NLP & RAG Search
:8003Multi-Agent RL Simulation
:8004React + TypeScript + Vite
:5173Dual-LLM Architecture: GPT-4o-mini for routing (10× cheaper) + GPT-4o for reasoning — reducing API costs by ~60%
Each service can be scaled, updated, or replaced independently — microservice independence with agentic orchestration.
Differential diagnosis engine based on the MEDDxAgent framework (NEC Research, ACL 2025). Generates ranked diagnoses with step-by-step clinical rationale using Chain-of-Thought prompting.
LSTM neural network predicting Emergency Department trolley counts across 12 Irish hospitals. Uses 7-day feature windows with autoregressive forecasting for 1–14 day predictions.
Fine-tuned BERT model (439MB) classifying ICD-10 procedures from diagnosis sequences. Sequential pathway prediction considers prior procedures for context-aware recommendations.
Semantic search engine over 1,203 clinical notes using Bio_ClinicalBERT embeddings and FAISS vector indexing. Captures medical context — not just keywords — for instant retrieval across discharge summaries, nursing notes, and radiology reports.
Multi-agent reinforcement learning for hospital resource optimization. Simulates 9 departments with 4 staff types, using both MADDPG and MAPPO algorithms with 5-stage curriculum learning.
Clinical operations dashboard with 6 interactive pages — from multi-hospital trolley comparison charts to a session-based agent chat with approval gates for diagnosis and treatment results.
The only platform combining all 6 clinical AI capabilities in a single deployable stack — with research-backed diagnosis, human-in-the-loop safety, and enterprise-grade infrastructure.
Single natural-language entry point for 6 AI services — eliminates the need for clinicians to learn 6 different UIs. One question, all capabilities.
Mandatory clinician approval for diagnosis and treatment recommendations. The AI never presents as final authority — clinical decisions are blurred with CSS until acknowledged.
PulseDiagAgent implements the MEDDxAgent framework (ACL 2025) — showing 10%+ accuracy improvement with Chain-of-Thought explainability.
GPT-4o-mini for routing (10× cheaper) + GPT-4o only for complex reasoning — reduces API costs by approximately 60%.
MIMIC-IV dataset provides realistic patient demographics, diagnoses, procedures, and clinical notes. 24,210 HSE trolley records from 12 Irish hospitals.
Docker, Kubernetes (GKE), CI/CD with GitHub Actions, Prometheus monitoring, Grafana dashboards, structured JSON logging — enterprise-grade infrastructure.
Explainable differential diagnosis in 5-10 seconds vs. manual review. 1-14 day ED trolley forecasting with breach risk indicators.
Both MADDPG and MAPPO for hospital optimization with 5-stage curriculum learning — enables comparison and selection per scenario.
Built on battle-tested open-source technologies — from LangGraph orchestration to Kubernetes deployment.
| Layer | Technology | Purpose |
|---|---|---|
| AI Orchestration | LangGraph ≥0.2.0 | Stateful agent workflow graph with 11 tools |
| LLM Provider | OpenAI GPT-4o / GPT-4o-mini | Reasoning, routing, diagnosis |
| ML Framework | PyTorch (Latest) | LSTM, BERT, MADDPG, MAPPO |
| NLP | HuggingFace Transformers | BERT, Bio_ClinicalBERT embeddings |
| Vector Search | FAISS | Semantic similarity search over clinical notes |
| API Framework | FastAPI / Flask | RESTful microservice APIs |
| Frontend | React 19 + TypeScript 5.9 | Clinical operations dashboard |
| Database | MongoDB 6.0 | Centralized data store |
| Monitoring | Prometheus + Grafana | Metrics scraping + dashboards |
| MLOps | MLflow v2.9.2 | Experiment tracking, model registry |
| Containers | Docker / Kubernetes (GKE) | Containerization & orchestration |
| Cloud | Google Cloud (GKE Standard Zonal) | europe-west1-b cluster deployment |
Every clinical decision passes through mandatory approval gates. The AI never presents as final authority.
Diagnosis and treatment results are blurred with CSS filter: blur(6px) and require clinician acknowledgment before viewing. The ApprovalGate component enforces this on every clinical response.
No secrets hardcoded in source code. OPENAI_API_KEY via environment variable only. Kubernetes secrets not checked into Git. .env.example template provided.
Maximum 15 iterations enforced as safety guard against infinite tool-call loops. Agent forced to validate regardless of pending calls.
High-contrast text, focus rings on all interactive elements, semantic HTML, and aria-hidden for decorative elements. Accessible to all clinicians.
Tool failures caught and returned as messages. MongoDB unavailability triggers graceful degradation. OpenAI rate limits handled with tenacity exponential backoff.
GDPR compliance checklist documented. PHI anonymization requirements. Data retention policies. MIMIC-IV requires PhysioNet credentialing + CITI training.
HSE Pulse is the only platform combining all 6 capabilities — diagnosis, forecasting, treatment, note search, resource optimization, and agentic orchestration — in a single deployable stack.
| Platform | Focus | HSE Pulse Differentiator |
|---|---|---|
| Epic Cognitive Computing | EHR-integrated CDS | Proprietary & tightly coupled to Epic EHR; HSE Pulse is EHR-agnostic |
| IBM Watson Health | Oncology treatment | Narrow cancer focus; HSE Pulse covers 6 clinical domains |
| Google MedPaLM | Medical Q&A | Research-only, no deployment tooling; HSE Pulse is production-ready |
| John Snow Labs | Clinical NLP | NLP-only; HSE Pulse integrates NLP + forecasting + RL + diagnosis |
| Nuance DAX / Microsoft | Ambient documentation | Dictation focus; HSE Pulse provides decision support |
| Viz.ai | Stroke detection | Imaging-only; HSE Pulse works with text-based clinical data |
HSE Pulse combines 6 specialized AI services under a single LLM-powered orchestrator — production-ready with Docker, Kubernetes, and enterprise-grade monitoring.