Skip to content
AI-native · 5 providers · self-throttling · every action reversible

AI-native infrastructure. Not "AI-powered" marketing copy.

Most platforms add a chatbot and call themselves AI-native. Sumeru treats AI as a first-class plane: inference rate-limited at the helper level, decisions orchestrated through typed events, every action reversible. Safety isn't documentation — it's enforced.

Request demo View architecture Read the platform →
— By the safety primitives

Enforced at the runtime. Not at the wiki.

4
AI surfaces
5
Providers routed
18
Trigger events
13
Action handlers
7-day
Mandatory dry-run
Every
Action reversible
01 The four surfaces

Inference. Decision. Recommendation. Autonomy.

Each surface has explicit primitives, observable behavior, and operational guarantees. Together they form the AI substrate every other module depends on.

01
Surface

Inference layer

Every AI call self-throttles. Routes can't bypass it.

callAI() is the canonical helper. It enforces per-shop rate limits, monthly budget caps, and provider fallback before issuing any request. Adding a new AI-powered route means you inherit the gates automatically — no per-route remember-to-call pattern.

Primitives
checkRateLimit(shop, feature)checkAiBudget(shop, estimatedTokens)Plan-tier-aware ceilingsTagged errors: RATE_LIMIT | BUDGET_EXCEEDED
02
Surface

Decision orchestration

Typed trigger events route through a dispatcher with snapshot, audit, and undo.

Detection paths emit autoaction events (e.g., seo.content_decay_detected). The dispatcher loads matching rules, evaluates triggerCondition against the event payload, snapshots the resource state, dispatches to action handlers via BullMQ, and writes a plain-language audit row. Every action is reversible.

Primitives
18 typed trigger events13 action handlersPer-rule cooldown + quiet hoursSnapshot-and-undo on every fire
03
Surface

Recommendation systems

Cohort-aware next-best-action wired to every customer channel.

The AI Recommendation Engine consumes Customer360 lifecycle stage, fatigue score, channel preference, and product affinity. Output: a ranked NBA list per customer with a confidence score. Wired to email, WhatsApp, SMS, and ad audience uploads.

Primitives
Lifecycle-aware NBASend-time optimization (pickOptimalSendTime)Bayesian discount-tier learnerA/B variant attribution
04
Surface

Autonomous workflows

DAG-driven flows with approval gates and dry-run safety.

Campaign orchestration plans, content calendars, and recovery sequences run as DAGs. 'Autonomous' doesn't mean unsupervised — the system proposes, the human approves (or auto-approves under merchant-set thresholds), the system executes. Mandatory 7-day dry-run for paid-spend actions.

Primitives
DAG plan executionApproval queues for agency-managed shops7-day mandatory dry-run for ads:bid:adjustPlain-language audit on every node
02 Position language

Words we ban. Words we earn.

Buyers in this category don't want to be sold to. They want to read the system. We ban the language that signals lazy thinking and replace it with words that map to runtime behavior.

Forbidden language

What we never say.

Lazy thinking. Each phrase signals we couldn't articulate the actual mechanism.

  • AI-powered
  • Smart [anything]
  • Magic
  • Revolutionary
  • Transformative
  • Next-generation
Earned language

What we say instead.

Specific, observable, bounded. Each phrase maps to runtime behavior with a citation.

  • AI-native
  • Inference layer
  • Decision orchestration
  • Autonomous workflows (with approval gates)
  • Predictive intelligence (with confidence intervals)
  • Recommendation system (named, sourced)
03 Worked example

What it looks like in production.

A real automation, end-to-end, with the actual primitives the system uses.

scenario · content decay → automated refresh reversible · audited · throttled
  1. 01

    Detect

    Decay detector runs on the GSC ingest. Compares trailing 30-day clicks to prior 30-day baseline. Identifies pages with ≥30% drop. Persists rows to SeoOpportunity.

  2. 02

    Emit

    After persistence, the detector emits one seo.content_decay_detected event per finding (capped at 50 per run). Payload carries page URL, drop %, and click delta.

  3. 03

    Dispatch

    Dispatcher loads matching rules. Per rule: evaluate triggerCondition (e.g., dropPct > 30), check daily cap, quiet hours, cooldown, freeze window. If clear, snapshot the SeoContent row.

  4. 04

    Act

    Action refresh_blog_post calls into the AI bulk queue. Worker self-throttles on shop AI budget. Generates updated content. Writes after snapshot. Status flips to applied.

  5. 05

    Audit + (optional) undo

    Plain-language audit row written to AutoActionExecution with before/after diffs. Notification fan-out (Slack/email/WA) per rule config. Merchant can undo from the timeline UI.

Ready to read the system instead of being sold to?

60-min architecture session with our engineering team. Whiteboard, runtime walk-through, Q&A on the specific primitives above.

Talk to engineering View architecture Developer docs →