precomputed-ai

Snapshots

The artifact refreshes on a named cadence. Like a newspaper, not a livestream.

Problem

The reasoning is stable enough to precompute, but the underlying world state changes — prices shift, models are released, vendor architectures evolve. A precomputed artifact that never regenerates becomes silently wrong. A live LLM call on every request is the wrong solution to staleness.

Mechanism

A scheduled job — daily, weekly, or triggered by an external event — regenerates the artifact. Between runs, the application serves from the static snapshot. The freshness window is surfaced to the user, typically as a “last refreshed” timestamp.

Required properties

Canonical example

CloudEstimate prices self-managed workloads across AWS, GCP, and Azure against scheduled regional pricing snapshots and serves LLM-generated sizing explanations from a versioned artifact cache. Both the pricing and explanation artifacts refresh daily via GitHub Actions. The “Pricing data last refreshed” footer surfaces the staleness window directly to the user.

When it fits

The world state changes on a knowable cadence. Users tolerate a known staleness window. Regeneration cost is bounded and predictable.

When it doesn’t

The world state changes faster than any reasonable refresh interval. Users require strict real-time accuracy. Regeneration cost approaches the saved inference cost.