One runtime. Four planes. Engineered for operators.
Sumeru Systems is structured as four independently scaling planes — data, intelligence, decision, action. Each plane has clear contracts, observable boundaries, and operational primitives that reflect how production commerce systems actually behave.
Independently scalable. Operationally fused.
The four planes.
Each plane is a distinct concern with its own scaling characteristics, ownership boundaries, and operational guarantees.
Data plane
Event ingest. Multi-tenant. Schema-isolated.
Every storefront event, every webhook, every API call lands in a tenant-scoped event stream. Idempotent ingest, replay-safe, and indexed for sub-second query.
Intelligence plane
Attribution, decay, anomaly, recommendation.
The intelligence plane runs on the data plane. Multi-touch attribution closes the loop in <60s. SEO decay detection runs on a 14-day baseline. Per-keyword performance Z-score anomaly detection runs on AdInsight rollups.
Decision plane
Typed trigger events → action registry → snapshot + undo.
The decision plane consumes signals from the intelligence plane. A dispatcher matches events to subscribed automations, evaluates triggerCondition, snapshots state, queues the action, and writes a plain-language audit row.
Action plane
Outbound to Shopify, ad platforms, messaging.
Every mutation goes through a self-throttling helper. Shopify GraphQL is cost-aware. AI calls are rate-limited + budgeted at the helper level. Ad-platform calls go through per-account token buckets.
Inference is a first-class plane, not a feature.
Every AI call routes through one helper that enforces per-shop rate limits, monthly budget caps, and provider-level fallback. Routes can't bypass it. Workers can't bypass it. New code can't bypass it — CI fails the PR.
Five providers. One interface.
- Anthropic Claude primary
- OpenAI GPT fallback
- Google Gemini multimodal
- Mistral budget
- Cohere classification
Self-throttling. Always-on.
- Per-shop rate limit — sliding window, plan-tier-aware
- Monthly budget cap — hard ceiling, alerts at 80% and 100%
- Tagged errors — code: "RATE_LIMIT" | "BUDGET_EXCEEDED" with retryAfterSec
- Usage logged — every call writes UsageLog with token counts + estimated cost
- CI gate — new code missing the gate fails the build
18 events. 13 actions. One audit log.
The decision plane subscribes to typed trigger events emitted by the intelligence plane. Every fire is captured as an AutoActionExecution row with snapshot, after-state, audit log, and undo handler.
Read the full architecture, or talk to engineering.
Public API docs, runtime guarantees, and the canonical event catalog — or a 60-min architecture session with our engineering team.