Architecture
Shipped LensAI pipeline vs design-target suite control plane. Deep links to eng DESIGN docs.
Today the runnable architecture is the LensAI streaming pipeline. The full observe → route → drift → retrain control plane is the product thesis — parts of it are still Scaffold or Design target. See Status & maturity.
Shipped — LensAI pipeline
Clients / eBPF tracer (mock or Linux)
|
| POST /ingest or /events
v
+---------------------+
| Rust ingest :8080 |
| WAL + rate limit |
+----------+----------+
|
v
Redpanda
|
v
+----------+----------+
| Go consumer |
| → ClickHouse |
+----------+----------+
|
v
Grafana :3000
distributed-flagd :8081 (etcd) — feature flags alongside the pathEngineering detail: ARCHITECTURE.md, DESIGN.md. Quickstart: lensai-integration.
Target — suite control plane
Design target / not shipped
Design target / not shipped
Agent / app / SDK
|
| POST /v1/chat/completions ← not shipped
v
+---------------------+
| Inferix :4000 | ← design target
| LensAI sample |
| TraceForge root |
| RouteIQ decide |
+----------+----------+
|
v
owned models / providers
|
v
DriftWatch → FineForge ← still ScaffoldDo not document :4000 as a running install. When implemented, Inferix sits next to gateways and existing traces — not instead of them.
Product roles
| Product | Maturity | Role |
|---|---|---|
| LensAI | Active | Latency, cost, errors, volume — ingest → warehouse → Grafana |
| TraceForge | Active (early) | Agent/tool/model spans — collector schema + streaming TraceForge packages |
| RouteIQ | Active (early) | Policy routing arc in infra-ai-streaming (cache → budget → decide → fallback) |
| DriftWatch | Scaffold | Async quality windows and alerts |
| FineForge | Scaffold | Retrain / promote / rollback jobs |
Eng design docs
- infra-ai-streaming DESIGN.md — CAP, partitioning, failure modes
- agent-trace-collector DESIGN.md — TraceForge span model
- API reference — shipped ingest/flagd vs target
:4000