Architecture

ACSI · Hospitality · 2020–Present · Public

PublicRepresentative · synthetic data
Live diagram — ingest → triage → language → product → issue → enrichment → draft → QC → route. Only draft + QC are metered; everything else is deterministic / local at $0.

One pipeline, nine stages

The engine is a single orchestration pipeline with a uniform stage contract (stage · provenance · model · tokens · cost · confidence). Stages 1–6 (ingest, triage, language, product, issue, enrichment) and stage 9 (route) are deterministic or local — $0 in both modes. Only stage 7 (draft) and stage 8 (QC) are metered.

The evidence is the looked-up facts: the draft step is constrained to those facts, dropping any sentence whose fact couldn't be resolved rather than fabricating it; the QC step claim-checks the draft against them.

Dual approach: cloud and OSS

The two metered stages are swap-ready. Cloud runs claude-haiku-4-5, cost-capped and fail-closed — at the session cap it stops and falls back to the $0 OSS path. OSS runs a small writer + checker recorded on local M4 hardware, replayed at $0. The deterministic stages run identically in both modes.

In this prototype everything is a deterministic mock at $0; the inspector labels stages 7 and 8 as the metered stages and shows simulated cost figures on the cloud/OSS toggle. The exact metered-vs-deterministic placement and the data model are the PRD's to confirm.

Architecture · Multilingual Email Orchestration · Abhishek Saxena