GOVERNED CLIENT OPERATIONS

Find the first real onboarding blocker. Propose the safest recovery.

Six bounded agents reconcile source coverage, classify the full journey, assess SLA risk, and produce an evidence-backed recovery proposal. Deterministic rules decide what failed. Human approval controls what happens next.

Run a scenario ↓ Inspect the API Try with your data

6: bounded agents
15: evaluation measures
17: trace dimensions
0: automatic mutations

01Recovery
CoordinatorLangGraph state

Sourcereliability

Journeyclassifier

SLArisk

Policyapproval gate

Every claim keeps its evidence hash

THE OPERATING PROBLEM

Paid clients can appear “broken” while they are still inside a legitimate grace window, or appear healthy while a source is unavailable.

THE SYSTEM DECISION

Classify six stages from verified facts, separate waiting from failure, stop on unknown data, and surface the earliest actionable blocker.

THE HARD BOUNDARY

The public system drafts a proposal only. A manager decision is recorded, but no CRM, email, community, or account tool is allowed to mutate.

LIVE SYNTHETIC WALKTHROUGH

Change the evidence. Watch the council change its decision.

Use a preset or change any source state. This is a real LangGraph run through the same typed API exposed in Postman and OpenAPI.

Council run

Waiting for evidence

Ready

↳

No run yet

Select a scenario, change its evidence, and run the council.

SYSTEM DESIGN

LLMs explain. Rules decide. Humans authorize.

The architecture separates stochastic reasoning from business truth and external side effects.

INPUT

FastAPI boundaryPydantic v2 validation
tenant and idempotency keys

→

PARALLEL ANALYSIS

LangGraph councilsource reliability
journey classification
SLA risk

→

CONTROL

Policy gateinterrupt and resume
named human decision

→

OUTPUT

Draft reportevidence references
no external mutation

Why this mechanism

FastAPI over Flask: The boundary needs generated OpenAPI, typed validation, and consistent error contracts without a custom validation layer.
LangGraph over free-form chat: Explicit state, parallel branches, checkpoints, step budgets, and interrupt-based approvals are inspectable and testable.
REST over GraphQL or gRPC: Operations teams can test every contract in a browser or Postman. The workflow has command-shaped routes, not a flexible query problem.
Deterministic classifier over LLM judgement: Grace windows and failure precedence are business policy. They must be reproducible, versioned, and covered by golden cases.

PRODUCTION ASSURANCE

Measured at the run, agent, tool, and decision levels.

Portfolio replay is zero-cost. A client deployment switches checkpoints and run storage to PostgreSQL and enables authenticated integrations.

Evaluation protocolLoading

Fifteen measures run across twelve golden scenarios.

schemataskstatusactionevidenceargumentsauthorizationapprovalidempotencyrecoverylatencystepscostinjectionreplay

Observability protocol17 dimensions

LangSmith traces the graph. OpenTelemetry exports vendor-neutral spans. Sentry captures application errors without default PII.

run → source_audit → classify → coordinate → interruptcontent.capture = false

Deployment boundary$0 demo

Docker Compose starts the full local system.
Vercel hosts the synthetic demonstration.
PostgreSQL enables durable production resume.
OpenRouter live narration is optional.
No paid fallback activates automatically.