Skip to main content

Zynomi Agent Fabric (ZAF)

Zynomi Agent Fabric (ZAF) — the platform a fleet of Zynomi agents runs on, in every department, grounded in a clinical knowledge engine built from a clinical ontology and GraphRAG, with every agent output carrying provenance and an accept/reject disposition.

In development

ZAF is the platform Zynomi is building through 2026 — the AI-native rebuild year. This page describes architecture and design intent, not shipped features. Nothing on this page is generally available yet.


Four architecture facts

FactWhat it means
Agent + human approvalEvery workflow pairs an agent with a human approval step. Agents do the data work; humans approve every result.
Ontology + GraphRAG groundingAgents are grounded in a clinical knowledge engine built from a designed clinical ontology and GraphRAG over a temporal knowledge graph — not free-form generation.
Semantic layer + MCP interfaceThe semantic layer (governed metrics) and the Model Context Protocol are the cross-system interface. Tools compute; the model narrates.
Provenance + dispositionEvery agent output carries provenance (what it was derived from) and an accept/reject disposition recorded by the reviewing human.

The third data layer

Clinical platforms already have two data views. ZAF adds a third:

LayerQuestion it answersStore
TransactionalWhat is happening now?OLTP (Frappe / PostgreSQL)
AnalyticalWhat happened, at scale?Lakehouse + semantic layer (dbt / Cube.dev)
Temporal knowledge graphWhat is connected, how did it evolve, and why?Graphiti → FalkorDB

The graph is built from the audit trail: business events flow through a change-data-capture (CDC) stream into Graphiti, which extracts and resolves entities and relationships into a FalkorDB temporal graph. Every edge carries valid-from / valid-to timestamps, so the graph can answer as-of questions — for example, an auditor's "explain everything around Amendment 4" in one traversal.

Designed ontology, not inferred

The ontology is two-tier and owned by the enterprise, populated by AI:

  • Tier 1 — first-class trial entities (study, site, subject, visit, form, query…)
  • Tier 2 — entities promoted from the graph: Protocol, Amendment, Change Request, Review, Approval, Deviation

CDISC SDTM/ODM exports are modeled as export-artifact nodes with derived-from provenance back to the source records.

A fleet of bounded agents

A supervisor agent owns no data tools. Bounded specialist agents each see exactly one engine, and are read-only by construction:

AgentEngine it sees
Trial Metrics AnalystSemantic layer (governed metrics)
Study HistorianTemporal knowledge graph
Operations AgentTransactional reads
Data StewardData quality and lineage

Agent runtime targets Amazon Bedrock — consistent with the platform's AWS-only hosting.