Architecting & Deploying DORA Songsmith on Google Cloud: A Zero-Trust, Multi-Agent Serverless Media Pipeline

DORA Songsmith is a multi-agent AI orchestration service designed to transform free-form DORA Community event descriptions into 3-minute, high-fidelity music videos. Built specifically for the DORA community discussions, DORA Songsmith pairs generative models, Gemini for lyric writing and DORA research synthesis, Lyria 3 Pro for music composition, and Veo for cinematic video generation with a robust, enterprise-grade Google Cloud deployment.

In this post, I will share the Google Cloud architecture powering DORA Songsmith, highlighting how Cloud Run, Secret Manager, and Terraform combine to deliver a zero-trust, supply-chain-secured multi-agent pipeline.

Architecture Overview: The Multi-Agent Pipeline

Songsmith delegates tasks across specialized serverless microservices coordinated by a central orchestrator.

DORA Songsmith Multi-Agent Architecture

Key Microservices

Lyricist Agent:
- Queries the latest DORA research (2025 metric benchmarks) via Gemini.
- Generates a structured 3-minute "Song Map" JSON containing timestamped lyrics (start_ms, end_ms), visual overlay triggers, and musical style cues.
- Persists output to songsmith-lyric-maps-${PROJECT_ID}.
Audio Engineer Agent:
- Translates the Lyricist's mood board into genre, tempo (BPM), and instrumental prompts for Google's Lyria 3 Pro music model.
- Outputs a 3-minute .wav track stored in songsmith-media-${PROJECT_ID}.
Video Producer Agent:
- Synthesizes an evolving 3-minute cinematic background video aligned with DORA themes using Veo.
- Writes the base .mp4 file to songsmith-media-${PROJECT_ID}.
FFmpeg Serverless Compositor:
- Deployed as a high-compute Cloud Run Job (4 vCPU, 8 GiB RAM, 30-minute timeout).
- Downloads raw audio, base video, and lyric timestamps, burning dynamic karaoke-style subtitle overlays into the final video file via FFmpeg.
Antigravity Orchestrator:
- Asynchronously coordinates parallel execution of Audio & Video generation, polls job status, and returns a time-bound V4 Signed URL to the client.
Web UI:
- React single-page application deployed to Cloud Run, protected by Identity-Aware Proxy (IAP) for secure authenticated access.

Serverless Compute on Google Cloud Run

Songsmith leverages Google Cloud Run (v2 API) for all backend execution, benefiting from:

Automatic Scaling & Scale-to-Zero: Eliminates idle infrastructure cost while handling bursty video synthesis workloads.
Service vs. Job Separation: Long-running heavy rendering operations are offloaded from HTTP APIs to Cloud Run Jobs, preventing HTTP timeout constraints.
Granular IAM Authentication: All inter-service calls use IAM authentication header tokens (roles/run.invoker) rather than exposed public endpoints.

# Excerpt from terraform/cloudrun.tf: Serverless Compositor Job
resource "google_cloud_run_v2_job" "compositor_job" {
  name     = "compositor-job"
  location = var.region

  template {
    template {
      timeout         = "1800s" # 30-minute timeout for rendering
      service_account = google_service_account.compositor_sa.email

      vpc_access {
        network_interfaces {
          network    = google_compute_network.songsmith_vpc.name
          subnetwork = google_compute_subnetwork.songsmith_subnet.name
        }
        egress = "ALL_TRAFFIC"
      }

      containers {
        image = "us-central1-docker.pkg.dev/${var.project_id}/songsmith-repo/compositor@sha256:..."
        resources {
          limits = {
            cpu    = "4"
            memory = "8192Mi"
          }
        }
      }
    }
  }
}

Zero-Trust Security Model & Keyless Runtime

The deployment enforces a strict Zero-Trust Security Policy:

1. Dedicated Service Accounts (Least Privilege)

Every component operates under a distinct Google Cloud Service Account with strictly bounded permissions:

Microservice	Service Account	Core IAM Roles Granted
Lyricist Agent	`lyricist-agent-sa`	`roles/aiplatform.user`, `roles/secretmanager.secretAccessor`, `roles/storage.objectUser`
Audio Engineer	`audio-engineer-sa`	`roles/aiplatform.user`, `roles/storage.objectUser`
Video Producer	`video-producer-sa`	`roles/aiplatform.user`, `roles/storage.objectUser`
Compositor Job	`compositor-sa`	`roles/storage.objectUser`
Orchestrator	`orchestrator-sa`	`roles/run.invoker`, `roles/run.developer`, `roles/iam.serviceAccountTokenCreator`

2. Secret Manager Integration

Zero API keys or service credentials are passed as environment variables or committed to source control. Runtime secrets are dynamically fetched from Secret Manager at startup using Cloud Run identity tokens.

3. Enforced Private Storage Buckets

All Cloud Storage buckets enforce Uniform Bucket-Level Access (uniform_bucket_level_access = true) and Public Access Prevention (public_access_prevention = "enforced"). Clients obtain temporary access to generated MP4 files strictly through cryptographic V4 Signed URLs generated by the Orchestrator via roles/iam.serviceAccountTokenCreator.

Software Supply Chain Security & Network Perimeter

Container Image Attestation via Binary Authorization

Songsmith implements end-to-end container security using Artifact Registry, Artifact Analysis On-Demand Scanning, and Binary Authorization:

Vulnerability Scanning: Cloud Build scans images for CVEs via gcloud artifacts docker images scan.
Cryptographic Attestation: Approved images are signed by songsmith-attestor using asymmetric keys in Cloud KMS.
Deployment Gatekeeping: Cloud Run blocks deployment of any image lacking a valid BinAuthz attestation signature.

# Excerpt from cloudbuild.yaml: Binary Authorization Attestation Step
- name: 'gcr.io/google.com/cloudsdktool/cloud-sdk'
  entrypoint: 'bash'
  args:
    - '-c'
    - |
      DIGEST=$(cat /workspace/digest.txt)
      gcloud beta container binauthz attestations sign-and-create \
        --project=${PROJECT_ID} \
        --artifact-url="${_REGION}-docker.pkg.dev/${PROJECT_ID}/${_REPO_NAME}/web-ui@$${DIGEST}" \
        --attestor="songsmith-attestor" \
        --attestor-project=${PROJECT_ID} \
        --keyversion-project=${PROJECT_ID} \
        --keyversion-location=${_REGION} \
        --keyversion-keyring="binauthz-keyring" \
        --keyversion-key="binauthz-key" \
        --keyversion="1"

Network Isolation with Custom VPC & Direct Egress

All Cloud Run services route traffic through a custom VPC and subnetwork via Serverless VPC Access. Direct VPC egress prevents rendering workers from making unauthenticated external network outbound requests.

Infrastructure as Code (IaC) with Terraform

The entire infrastructure lifecycle is declared inside terraform/, structured into modular definitions:

cloudrun.tf: Configures Cloud Run Services, Cloud Run Jobs, ingress rules, and BinAuthz enforcement.
iam.tf: Defines dedicated Service Accounts and precise IAM role bindings.
storage.tf: Manages encrypted GCS buckets and IAM object user policies.
network.tf: Sets up VPC networks, subnets, and firewall configurations.
secrets.tf: Provisions Secret Manager resources.
cloudbuild.tf: Configures Cloud Build triggers and IAM permissions.

Operational deployment helper scripts in deploy/ (e.g. deploy_updates.sh, deploy_orchestrator.sh) automate image building, vulnerability scanning, BinAuthz signing, image digest injection, and terraform apply.

Conclusion

DORA Songsmith demonstrates how modern AI multi-agent systems can be operationalized safely on Google Cloud. By anchoring agent workflows in Cloud Run, protecting secrets with Secret Manager, enforcing Binary Authorization attestations, and managing infrastructure through Terraform, Songsmith provides a scalable template for enterprise-grade generative media pipelines.