Reference architecture · Banking

AI-enabled Cash Management System on Tencent Cloud

A multi-AZ, multi-VPC architecture centered on secure transaction processing, controlled AI assistance, and end-to-end observability. The diagrams emphasize traffic flow, trust boundaries, and monitoring coverage.

Flow 01

Production request and data flow

Public traffic is protected at the edge, distributed across two isolated application VPCs, and routed to multi-AZ TKE clusters. Shared data services reside in a private data VPC.

Separate public and private entry

Internet users enter through EdgeOne. Bank data centers and internal channels use Direct Connect or VPN and internal CLB, avoiding an unnecessary public path.

Edge protection is not private connectivity

EdgeOne provides DDoS, WAF, Bot protection, TLS and origin protection. It does not replace private routing for core banking integration.

Multi-VPC means two TKE clusters

Each TKE cluster belongs to one VPC. VPC A and VPC B therefore operate independent clusters; each cluster spans two availability zones.

AI Agent is an application workload

Deploy Agent replicas in both application VPCs. Keep prompts, planning state and tool execution inside the controlled application boundary.

No direct database access for AI

The Agent calls allowlisted CMS APIs through a policy and audit gateway. Payment, transfer and limit changes require explicit approval.

TokenHub centralizes model access

One OpenAI-compatible endpoint provides model routing, API-key governance, quota isolation and standardized usage monitoring.

Private data VPC

Redis stores reconstructable cache and coordination state. TDSQL-D remains authoritative for balances, payments, ledgers and limit utilization.

2Independent application VPCs

2 × 2TKE clusters × availability zones

1Private data VPC and system of record

0Direct AI-to-database connections

Flow 02

RUM, APM and AI Agent observability

A single trace context connects the user experience to application spans, model calls and governed tools. Infrastructure products retain their own metrics and logs.

Application traceAI / Agent traceProduct metrics and logs

RUM starts at the browser

The RUM SDK measures real user experience and injects traceparent into selected API calls. It does not expose internal EdgeOne or CLB processing spans.

APM covers the application path

OpenTelemetry follows the request from an instrumented Ingress or CMS API through service-to-service, Redis-client and SQL-client spans.

Dedicated AI Agent telemetry

TCOP LLM/Agent observability records Agent entry, model metadata, input/output token counts, model latency, errors and nested tool calls.

Collectors in both TKE clusters

Instrument application and OpenAI-compatible calls, export through local OTel Collectors, and preserve the same Trace ID across model and tool calls.

TokenHub adds platform-level metrics

Monitor Tokens, TPM, RPM, TTFT, TPOT, error rates, rate limits and cache hit ratio by endpoint, model and API key, with TCOP alarms.

Control plane

Banking-grade AI controls

The Agent is useful only when its authority, data boundary and telemetry are deliberately constrained.

Access

Tool-level authorization

Expose only allowlisted CMS APIs.
Use scoped service identity and short-lived credentials.
Require approval for payments, transfers and limit changes.

Data

System-of-record protection

No direct Agent connection to TDSQL-D.
Redis remains reconstructable, non-authoritative state.
Validate every tool input at the gateway.

Observability

Metadata-first tracing

Set TRACELOOP_TRACE_CONTENT=false.
Do not export prompts, completions or secrets.
Retain model, token, latency, error and tool status.

Monitor end to end

Browser experience → API latency → service topology → Agent execution → model call → tool call → CMS API → Redis / SQL client latency.

Do not overclaim

OpenTelemetry cannot automatically reveal internal EdgeOne or CLB processing. Use cloud-product metrics, access logs and security logs for those layers.