メインコンテンツまでスキップ

AI Observability Demo

Multi-cloud AI-native full-stack observability platform for monitoring LLM workloads.

Quick Start​

Prerequisites​

  • AWS Account with Bedrock access. This demo uses Claude 3 Haiku/Sonnet in us-east-1, but you can substitute any Bedrock-supported model by updating the model ID in gateway/litellm-config.yaml. The observability pipeline is model-agnostic and works with any LLM provider supported by LiteLLM.
  • AWS CLI configured with AdministratorAccess
  • Docker Desktop running
  • Docker Compose v2
  • Python 3.11+
  • Terraform 1.5.0+

Phase 1: Infrastructure Provisioning​

cd AI-OBS_DEMO/terraform
terraform init
terraform plan -out=tfplan
terraform apply tfplan

# Capture outputs
export AMP_WORKSPACE_ID=$(terraform output -raw amp_workspace_id)
export AMP_REMOTE_WRITE_URL=$(terraform output -raw amp_remote_write_url)
export AMP_ENDPOINT=$(terraform output -raw amp_endpoint)

Phase 2: Environment Configuration​

cd ..
cp .env.example .env

# Edit .env with your AWS credentials and Terraform outputs
# Add:
# - AWS_ACCESS_KEY_ID
# - AWS_SECRET_ACCESS_KEY
# - AWS_SESSION_TOKEN (if using temporary credentials)
# - AMP_REMOTE_WRITE_URL (from Terraform)
# - AMP_ENDPOINT (from Terraform)

Phase 3: Build and Launch​

# Build all services
docker compose build

# Start the stack
docker compose up -d

# Verify services
docker compose ps

# Check OTEL Collector health
curl http://localhost:13133

# Run demo and watch logs
docker compose logs -f ai-app

Phase 4: Validate Telemetry​

# Install awscurl
pip3 install awscurl

# Query AMP for token usage
awscurl --service aps --region us-east-1 \
  "${AMP_ENDPOINT}api/v1/query?query=gen_ai_usage_input_tokens_total"

# Check X-Ray traces in AWS Console
# Navigate to: X-Ray > Traces > Filter by service: ai-observability-demo

# Check CloudWatch logs
# Navigate to: CloudWatch > Log Groups > /ai-observability-demo

Phase 5: Grafana Dashboard​

  1. Open AWS Console → Amazon Managed Grafana → ai-observability-demo
  2. Click "Open Grafana"
  3. Add data sources:
    • Prometheus: Use AMP endpoint with SigV4 auth
    • CloudWatch: Set to the region where your resources are deployed
    • X-Ray: Set to the same region as CloudWatch
  4. Import dashboard: grafana/dashboards/ai-observability.json

Phase 6: MCP Server Integration​

# Install MCP server dependencies
cd mcp-server
pip3 install -r requirements.txt

# Configure Kiro MCP
# Add to .kiro/mcp.json:
{
  "mcpServers": {
    "ai-observability": {
      "command": "python3",
      "args": ["/path/to/AI-OBS_DEMO/mcp-server/prometheus_mcp_server.py"],
      "env": {
        "AMP_ENDPOINT": "your-amp-endpoint",
        "AWS_REGION": "your-aws-region",
        "AWS_ACCESS_KEY_ID": "your-key",
        "AWS_SECRET_ACCESS_KEY": "your-secret",
        "AWS_SESSION_TOKEN": "your-token"
      }
    }
  }
}
Region Configuration

Set AWS_REGION to the region where your observability resources (AMP, CloudWatch, X-Ray) are deployed. Any AWS region that supports these services will work.

Natural Language Queries in Kiro​

Once configured, ask Kiro:

  • "Which model is consuming the most tokens right now?"
  • "What is the P95 latency for Claude Haiku over the last hour?"
  • "Have there been any throttle events in the last 30 minutes?"
  • "Estimate the cost of LLM usage today"
  • "Compare latency across all active models"

Teardown​

# Stop and remove containers
docker compose down

# Destroy AWS resources
cd terraform
terraform destroy -auto-approve

# Remove project (optional)
cd ../..
rm -rf AI-OBS_DEMO

Architecture​

  • OpenTelemetry: Vendor-neutral instrumentation with GenAI semantic conventions
  • LiteLLM: Multi-provider AI gateway (AWS Bedrock, Azure OpenAI, etc.)
  • Amazon Managed Prometheus: Scalable metrics storage with SigV4 auth
  • AWS X-Ray: Distributed tracing for LLM invocations
  • Amazon Managed Grafana: Unified visualization across AMP, CloudWatch, X-Ray
  • Custom MCP Server: Natural language observability queries via Kiro

Extending to Multi-Cloud​

  1. Azure OpenAI: Add real credentials to .env - LiteLLM handles routing automatically
  2. GCP Vertex AI: Add Vertex AI model to gateway/litellm-config.yaml
  3. On-premises models: Deploy vLLM/Ollama locally, add as LiteLLM provider
  4. Federated metrics: Deploy OTEL Collector in each cloud, remote write to same AMP workspace

References​