The pitch on decklar.com is clear: "the world's only Real-Time Decision AI Platform for Supply Chains, with Unified Visibility." Four of the Global Top 10 in Pharma. Six in CPG. Four in Automotive. Seven verticals where Decision AI decides at the edge — and every sensor check-in fires multiple inference calls in seconds, under regulated supply-chain data.
That scale isn't an emulator problem — it's an inference economics problem. Decision AI, Unified Visibility, and TrackerOS each call models on every sensor event, every shipment exception, every threshold trigger. The unit economics of "a decision in seconds" depend entirely on what's running underneath.
Three primitives that map directly to how Decklar's platform actually runs:
AI Gateway + Workers AI — under every Decision AI call: per-customer spend caps, fallback routing when a provider lags, one unified log across every model in the inference pipeline. The answer to "what's our cost-per-shipment of decisioning?"
Workers + Durable Objects — for the edge sensor triage layer: TrackerOS check-ins land at the nearest of 330+ POPs, get triaged in milliseconds, and only escalate to the central AI Platform when the decision actually needs it.
R2 + Vectorize — for the 10+ years of self-enriching visibility data: zero-egress historical archive plus semantic retrieval over decade-deep context for every new Decision AI inference.
Is the bigger near-term challenge on the economics side — keeping per-shipment Decision AI cost predictable as customer count grows — or on the scale side — moving more triage to the edge so the central platform isn't the bottleneck at 1,000 customers? 20 minutes to find the right starting point.
The detailed primitive-by-primitive mapping — including the eight things Cloudflare changes for Decklar, the AI Platform components × Cloudflare primitives table, the request-flow diagram for a single sensor check-in, and the 1,000-customer scale economics — runs about 47 KB of dense technical content.
Read the expanded version →