GRAPH/OS
A multi-tenant business-graph platform. Versioned facts, scoped agents, and nothing sends without a human.
What it is
A business-graph platform for built-to-order and e-commerce SMBs: businesses whose reality lives across email, order forms, commerce, and payments rather than a tidy CRM. GRAPH/OS maintains a versioned entity graph of the whole business and runs scoped agents over it.
The graph
- Entities, edges, and bi-temporal facts, so the graph answers both "what is true now" and "what did we believe then"
- pgvector entity resolution for deduping records that arrive from different sources under different names
- Source adapters normalize email, forms, commerce, and payment data into one queryable graph; the engine only ever sees normalized entities
The agents
- Scoped LLM agents produce draft-first actions: quotes, customer replies, order-risk flags, repricing
- Every action cites the graph nodes it traversed; cite-or-reject is enforced at runtime, so no agent acts on an ungrounded claim
- A human approves before anything sends; actions graduate to more autonomy only per an explicit approval policy
- Groq Llama handles extraction and classification, Claude Sonnet runs the agent reasoning loops
Multi-tenancy done properly
- Postgres row-level security per tenant from day one, enforced with a dedicated app role that cannot bypass it
- Tenant packs are configuration, never code: vocabulary, entity extensions, adapter bindings, and approval policy are per-tenant data
Operator console
React console with a review queue, per-run trace viewer, and graph browser, traced end to end with Langfuse and OpenTelemetry.
Stack
TypeScript monorepo. Vite React SPA, Express API, Neon Postgres with pgvector, Drizzle with hand-written idempotent SQL migrations.