Staging analyzers so a conversation can carry three risks at once.
A marketplace conversation can be abusive, fraudulent, and a credential lure simultaneously. This architecture runs a cheap primary screen, then fires specialized analyzers conditionally - each risk routed to the workflow that owns it.
Illustrative reference architecture. The scenario, volumes, and economics below are modeled - not a named customer case study. Validate every number against your own workload.
The scenario
A mid-size marketplace processes buyer–seller messaging plus public listings. A single flat classifier cannot tell "this message is rude" from "this message is a payment-diversion scam." The trust team needs each risk type routed differently: moderation for abuse, a transaction hold for fraud, and a security workflow for credential lures.
Before / after architecture
Before
With Sentiqra
Decision ownership stays with you
- The signal router, thresholds, and holds are yours - Sentiqra produces evidence, your decision engine acts.
- Each domain keeps its own queue and its own audit trail: moderation actions, transaction holds, security escalations.
- Analyzer selection is conditional - you pay for runs that fire, not for every analyzer on every message.
- The gate also decides what specialists never see: end-to-end recall is gate recall x specialist recall conditional on passing. Validate the frozen routing graph against a baseline that runs each specialist directly (sentiqra/workflow/v1) before adopting conditional routing.
Illustrative economics
Both columns are modeled from explicit assumptions. Do not treat these as guaranteed savings - model them against your own workload.
Current approach (modeled)
- N messages × flat classifier cost
- + manual split of flagged items into abuse / fraud / security buckets
- + wrong-queue cost: fraud treated as rudeness, lures as spam
With Sentiqra (modeled)
- N messages × primary screen (2 analyzers)
- + N × conditional rate × deeper analyzers
- = cheaper than N × all analyzers, with per-domain routing included
Example: 8M messages at 1.15 average analyzers ≈ 9.2M runs vs 32M with unconditional pipelines. At Production tier ($799/mo incl. 500K runs, $1.50/1K overage): 9.2M ≈ $13,849/mo (~$166K/yr) vs 32M ≈ $48,049/mo - a 3.5× cost difference from routing alone. Model your real conditional rate. See current pricing.
Failure behavior
Enterprise architects care about what happens when analysis is unavailable almost as much as accuracy. Define these states explicitly.
Sentiqra available
Normal staged flow - primary screen, conditional escalation, domain-owned routing.
One analyzer fails
Partial evidence - other domain signals still route; failed branch retries or holds for review.
Sentiqra unavailable
Customer fallback policy decides: allow, queue, or hold. Defined in your runbook, not improvised.
Quota / spend cap reached
Explicit state. Primary screen can keep running while conditional branches degrade gracefully.
Transaction-hold backlog
Holds age out per your policy - the signals that triggered them remain stored and replayable.
Illustrative operational impact
Three risks stay visible
Abuse, fraud, and lures in the same conversation each produce their own signal set - nothing collapses into one score.
Domain ownership
Moderation sees abuse signals, fraud ops sees transaction-risk signals, security sees lures. Each team gets its own queue.
Staged economics
The cheap primary screen runs on everything. Deeper analysis fires only when the primary screen warrants it.
What to measure in your evaluation
Run both approaches against the same labeled production sample. Compare:
The right migration decision should come from your workload - not a generic savings percentage.
Build the reference architecture
Start with 5,000 free analyzer runs and evaluate Sentiqra against your current pipeline.
No card required.