Get started

Architecture

Agents and apps send inference through Inferix. The control plane records metrics and spans, applies RouteIQ policy, forwards to an owned model or provider, then feeds DriftWatch and FineForge.

Request path

  Agent / app / SDK
           |
           |  POST /v1/chat/completions
           v
  +---------------------+
  |  Inferix :4000      |
  |  auth (master key)  |
  |  LensAI sample      |
  |  TraceForge root    |
  |  RouteIQ decide     |
  +----------+----------+
             |
             |  invoke chosen model
             v
  +----------+-----------+
  | owned/* endpoints    |
  | provider APIs        |
  +----------+-----------+
             |
             |  response + usage
             v
  LensAI metrics  ·  TraceForge spans
             |
             v
  DriftWatch windows  →  FineForge jobs
                         (promote / rollback)

Inferix sits next to gateways and existing traces. Point clients at :4000; keep your gateway for auth federation or rate limits if you already have one.

Product roles

Product	Where it sits	Operator output
LensAI	On every call ingress/egress	Latency, cost, errors, volume by model/agent/tenant
TraceForge	Around route + tool + model spans	Debuggable traces for a single user request
RouteIQ	Before model invoke	Policy decision: model name + fallback chain
DriftWatch	Async over golden / live windows	Alerts when quality score drops
FineForge	Job runner after drift (or schedule)	Train → evaluate → promote or rollback

Data stores (high level)

Config — inferix.yaml (models, routing, observe, drift). Hot-reload where supported; see Configuration.
Metrics buffer — LensAI time series (in-process + optional remote write).
Span store — TraceForge traces in local volume or Postgres; optional OTLP export.
Drift state — golden sets, window scores, alert history.
FineForge artifacts — job logs, candidate revisions, promote history.

Failure domains

Design for partial failure

If an owned model is down, RouteIQ walks the fallback chain and LensAI records the error class.
If TraceForge storage is slow, spans buffer locally; completions still return.
DriftWatch and FineForge are off the hot path — a stuck job does not block inference.

Full inferix.yaml reference.

Configuration →