[AI-OPS / Applied AI]

AI Customer Operations

Turns high-friction service requests into guided actions without giving the model unchecked control.

Agent-assisted customer operations built around real approvals, policy, and operational APIs.

Status
[active]
My role
Architecture / backend / agent design / frontend demo / evals / CI
Evidence model
Tests / CI / runtime proof

01

Request

Customer intent

02

Context

Travel Ops API

03

Proposal

Bounded AI

04

Approval

Human decision

05

Execute

Idempotent write

Live system traceMedia capture pending

[LIVE DEMO]

Open the Resolution Workspace.

Verify the sample booking, ask about the disruption, and inspect sources or the system trace.

Open live demo
[VERIFIED TRACE]Recorded execution

See where AI stops and deterministic software takes over.

Replay booking MX789C to follow how context, policy, approval, and idempotent execution move through the system.

Evidence
CI
35790449089
Comp smoke
35694512560
Mutation authority
Application
Step 01

Customer asks for help

The model receives intent, not authority. Trusted booking facts still have to come from Travel Ops.

Input
What can I do about booking MX789C?
Channel
Web interaction
Mutation
None
Execution

Not exposed as an LLM tool

Authority

Travel Ops owns facts + policy

Approval

Human state, fingerprint-bound

[01 / Business impact]

What changes for the business.

One workflow brings booking context, recovery choices, and compensation guidance together.

Execution stays approval-gated and replay-safe, reducing the blast radius of automation.

Operational policy—not the model—controls monetary outcomes and business rules.

[02 / PROOF]

What is already verified.

    017/7 foundation + 6/6 security/failure eval scenarios; agent-turn evals remain zero-mutation.
    02Private AI Ops API on Railway with verified booking access and scoped sessions.
    0364+ automated tests plus deterministic agent, security/failure, and retrieval eval gates.
    04Approved rebook flow proved bounded execution and replay idempotency.
    05Compensation smoke proved authoritative policy evaluation, execution, replay, and single persistence.

[03 / UNDER THE HOOD]

Architecture, controls, and implementation.

Technical detail lives here so the business story stays readable first.

Architecture diagram showing a traveler using AI Customer Ops Web, a Next.js and FastAPI application boundary, PydanticAI orchestration, trusted Travel Ops and Knowledge services, controlled outputs, and a separate human confirmation gate before execution.
Current connected architectureDashed path = next execution slice

My role

Architecture / backend / agent design / frontend demo / evals / CI

Technology surface

PythonPydanticAINext.jsFastAPIPostgreSQLpgvectorPydantic EvalsGitHub ActionsRailway

Next slice

Add customer confirmation for approved actions, then deepen retrieval calibration, tracing, and live-model eval coverage.

System index