KLH Architecture — Tier 1 vs Tier 2

Same journey, same persona, same language. Left returns from precomputed cache in <50ms at ~$0. Right forces live generation through Claude → shows real cost and latency. The architecture difference, demonstrated.

Journey
Persona
Verbosity
Tier 1 — Cache
Precomputed · Upstash Redis
⚡
—
Latency
—
Cost
—
Source
Select a journey and click Run
Waiting
Tier 2 — Live
Claude Sonnet 4.5 → Sarvam
🧠
—
Latency
—
Cost
—
Source
Select a journey and click Run
Waiting