10장: 실전 프로젝트 — AI-Native 시스템 아키텍처 설계

interface IncomingRequest {
  tenantId: string;
  userId: string;
  sessionId: string;
  message: string;
  metadata: {
    channel: "web" | "mobile" | "api";
    language: string;
    timestamp: string;
  };
}
 
interface ProcessedRequest extends IncomingRequest {
  tenantConfig: TenantConfig;
  routingDecision: RoutingDecision;
  traceId: string;
}
 
async function handleRequest(req: IncomingRequest): Promise<ProcessedRequest> {
  const traceId = crypto.randomUUID();
 
  // 1. 테넌트 구성 로드
  const tenantConfig = await tenantConfigService.getConfig(req.tenantId);
 
  // 2. 할당량 검증
  const quotaCheck = await tenantConfigService.enforceQuota(req.tenantId, {
    tokens: estimateTokens(req.message),
  });
  if (!quotaCheck.allowed) {
    throw new QuotaExceededError(quotaCheck.reason);
  }
 
  // 3. 모델 라우팅 결정
  const routingDecision = routeToModel({
    taskType: classifyIntent(req.message),
    inputLength: req.message.length,
    userTier: tenantConfig.tier,
    qualityRequired: tenantConfig.tier === "enterprise" ? "high" : "medium",
    latencyBudgetMs: 5000,
  });
 
  return {
    ...req,
    tenantConfig,
    routingDecision,
    traceId,
  };
}
 
function estimateTokens(text: string): number {
  // 한글 기준 대략적인 토큰 추정 (1 토큰 = 약 1.5 글자)
  return Math.ceil(text.length / 1.5);
}
 
function classifyIntent(message: string): string {
  // 빠른 규칙 기반 분류 (LLM 호출 전)
  if (message.includes("환불") || message.includes("취소")) return "refund";
  if (message.includes("주문") || message.includes("배송")) return "order-tracking";
  if (message.includes("사용법") || message.includes("어떻게")) return "how-to";
  return "general";
}

# 추론 워커 Deployment
apiVersion: apps/v1
kind: Deployment
metadata:
  name: inference-worker
  labels:
    app: ai-cs-platform
    component: inference
spec:
  replicas: 4
  strategy:
    type: RollingUpdate
    rollingUpdate:
      maxSurge: 2
      maxUnavailable: 0
  template:
    spec:
      terminationGracePeriodSeconds: 120
      containers:
        - name: inference-worker
          image: ai-cs-platform/inference-worker:latest
          resources:
            requests:
              cpu: "500m"
              memory: "1Gi"
            limits:
              cpu: "2000m"
              memory: "4Gi"
          readinessProbe:
            httpGet:
              path: /health/ready
              port: 8080
            initialDelaySeconds: 5
            periodSeconds: 10
          livenessProbe:
            httpGet:
              path: /health/live
              port: 8080
            initialDelaySeconds: 15
            periodSeconds: 20
          env:
            - name: REDIS_URL
              valueFrom:
                secretKeyRef:
                  name: ai-cs-secrets
                  key: redis-url
            - name: ANTHROPIC_API_KEY
              valueFrom:
                secretKeyRef:
                  name: ai-cs-secrets
                  key: anthropic-api-key