Guides

Platform operators

You own inference cost, quality, and agent reliability for the org. Inferix is the control plane in front of owned models and providers — observe, route, detect drift, retrain.

Scenario

Multiple agent teams ship prompts weekly. Finance wants $/successful task. SRE wants p95 and error budgets. You need one plane that attributes cost, enforces routes, and gates promotes — without replacing the gateway or APM.

What Inferix gives you

LensAI fleet SLOs + tenant cost on Grafana :3000
TraceForge RCA next to existing OTel
RouteIQ budgets, cache, cheap/general/hard matrix
DriftWatch slice alerts → FineForge canary promote

Setup

Bring-up checklist:

# 1. Deploy control plane :4000 + Postgres
# 2. Mount inferix.yaml with models + routing + observe + drift
# 3. Point Grafana at ClickHouse (LensAI) — port 3000
# 4. Require tenant_id / agent_id / trace_id at ingest
# 5. Wire FineForge promote to emit deploy markers
# 6. Page on driftwatch.alerts.open for critical agents

Install: Install & self-host. Config: Configuration.

Dashboards to watch

Panel	Signal	Action
Fleet p95 / error rate	LensAI by model_id	Failover / scale owned pool
Cost by tenant	lensai.cost.usd	Budget deny or cache-only
Route mix	routeiq.model.mix	Stop over-escalation
Open drift alerts	driftwatch.alerts.open	FineForge job
Canary score	shadow vs prod	Promote or rollback
Ingest lag	lensai.ingest.lag_ms	Scale consumers

Failure modes

Cardinality blowup (L01) — drop prompt text labels
Unattributed cost (L08) — reject missing tenant_id
Alert fatigue (D22) — critical slices only
Confounded deploys (D19) — change one variable per promote

Ownership

Platform owns Inferix + model registry pins. Agent teams own golden sets and business-success predicates. FineForge CI gates block bad prompt merges.

Order, policy, refund reference workload.

Support & CX agents →