MCMHUB BACKEND

CDSS platform — multi-agent virtual consilium for physicians reviewing complex cases with long medical history

VIRTUAL CONSILIUM

MULTI-AGENT

Parallel analysis by ReAct agents across medical specialties.

CHAIN-OF-EVIDENCE

RED FLAGS

Every statement is tied to a chunk or document. Facts are separated from hypotheses.

STACK

NestJSLangGraphLangChainKafkaPostgreSQLQDrantS3

ARCHITECTURE & PRODUCT INTENT

MCM (Medical Multiagent) is a CDSS backend for qualified physicians. It unifies medical documents, semantic search, and a team of specialized AI agents. The physician remains the attending doctor — the final decision stays human.

Composer → Doctor Agents → Supervisor: asynchronous pipeline via Kafka. Transactional outbox guarantees event delivery.
RLM + ToolCall: agents request data through search_patient_evidence, get_attachment_excerpt, get_record_fields — they do not invent lab results.
State machine: pending → in_queue → on_resolving → resolved/failed — transitions are validated, races excluded.

SAFETY & CLINICAL LOGIC

CDSS PRINCIPLE

The system does not make a final diagnosis — it forms a differential list with arguments, red flags, and follow-up recommendations. Not for patient self-diagnosis.

DOCUMENT PROCESSING

PDF, DOCX, OCR from images → chunking (1,500 characters, overlap 200) → embeddings in QDrant. Extracted text in S3, not PostgreSQL.

MCM CONSILIUM CONSOLE

Enter a patient complaint to simulate the consilium (for example, "persistent dry cough, low-grade fever" or "chest pressure after running"):

[ AGENT ACTIVITY LOGS ]:

Waiting for patient medical input.