GCP reference architecture.
Infrastructure as code.

resource "google_cloud_run_v2_service" "leave_agent" { name = "leave-agent" location = var.region # asia-south1 ingress = "INGRESS_TRAFFIC_INTERNAL_ONLY" template { service_account = google_service_account.leave_agent_sa.email scaling { min_instance_count = 0 max_instance_count = 10 } containers { image = "${var.region}-docker.pkg.dev/${var.project_id}/autohr-services/leave-agent:${var.image_tag}" resources { limits = { cpu = "2" memory = "1Gi" } } # Secrets from Secret Manager — no env var credentials env { name = "GEMINI_API_KEY" value_source { secret_key_ref { secret = google_secret_manager_secret.gemini_key.secret_id version = "latest" } } } env { name = "GEMINI_MODEL" value = "gemini-1.5-flash-002" # pinned — not latest } env { name = "RAG_CONFIDENCE_THRESHOLD" value = "0.80" } startup_probe { http_get { path = "/health" } initial_delay_seconds = 10 failure_threshold = 3 } } } }

# Firestore security rules — audit_log is append-only resource "google_firestore_document" "security_rules" { project = var.project_id collection = "_security_rules" document_id = "rules" # Inline rules — enforced at Firestore layer # No service account can bypass these } # The actual Firestore rules (firestore.rules) # deployed via firebase-tools in CI/CD: /* rules_version = '2'; service cloud.firestore { match /databases/{database}/documents { // Audit log: service accounts may CREATE only // UPDATE and DELETE are denied for ALL identities match /audit_log/{entry} { allow create: if request.auth != null; allow read: if request.auth != null; allow update: if false; // immutable allow delete: if false; // immutable } // Leave requests: agents may create and read // Workers may read their own records only match /leave_requests/{id} { allow create: if request.auth != null; allow read: if request.auth.uid == resource.data.employee_id || request.auth.token.role == "agent"; allow update: if request.auth.token.role == "agent"; allow delete: if false; } // Sessions: 24h TTL enforced via scheduled Cloud Function match /sessions/{id} { allow read, write: if request.auth.token.role == "agent"; } } } */

# Each service gets its own SA — minimum permissions resource "google_service_account" "leave_agent_sa" { account_id = "leave-agent-sa" display_name = "Leave Agent Service Account" } # Firestore: read + write (not admin) resource "google_project_iam_member" "leave_agent_firestore" { project = var.project_id role = "roles/datastore.user" member = "serviceAccount:${google_service_account.leave_agent_sa.email}" } # Secret Manager: access to leave-agent secrets only resource "google_secret_manager_secret_iam_member" "leave_agent_gemini" { secret_id = google_secret_manager_secret.gemini_key.id role = "roles/secretmanager.secretAccessor" member = "serviceAccount:${google_service_account.leave_agent_sa.email}" } # Pub/Sub: publish to AGENT_EVENT topic only resource "google_pubsub_topic_iam_member" "leave_agent_pubsub" { topic = google_pubsub_topic.agent_event.name role = "roles/pubsub.publisher" member = "serviceAccount:${google_service_account.leave_agent_sa.email}" } # Vertex AI: invoke Gemini models only resource "google_project_iam_member" "leave_agent_vertex" { project = var.project_id role = "roles/aiplatform.user" member = "serviceAccount:${google_service_account.leave_agent_sa.email}" } # NOTE: STT service account does NOT get Vertex AI role # Pub/Sub service account does NOT get Firestore role # No cross-domain access — enforced at IAM layer

resource "google_cloud_run_v2_service" "stt_service" { name = "stt-service" location = var.region ingress = "INGRESS_TRAFFIC_INTERNAL_ONLY" template { service_account = google_service_account.stt_sa.email scaling { min_instance_count = 1 # warm — eliminates GPU cold start max_instance_count = 5 } node_selector { accelerator = "nvidia-tesla-t4" } containers { image = "${var.region}-docker.pkg.dev/${var.project_id}/autohr-models/whisper-large-v3:${var.whisper_tag}" resources { limits = { cpu = "4" memory = "16Gi" "nvidia.com/gpu" = "1" } startup_cpu_boost = true } env { name = "WHISPER_MODEL" value = "large-v3" } env { name = "FALLBACK_TO_GOOGLE_STT" value = "true" # automatic fallback on error } env { name = "MIN_CONFIDENCE_THRESHOLD" value = "0.85" } } } }

# Five topics — one per pipeline stage locals { topics = [ "inbound-msg", "stt-result", "agent-event", "hitl-escalation", "notification-out", ] } resource "google_pubsub_topic" "topics" { for_each = toset(local.topics) name = "autohr-${each.key}" # Retain messages 7 days for replay on failure message_retention_duration = "604800s" } # Dead-letter topic for all critical subscriptions resource "google_pubsub_topic" "dead_letter" { name = "autohr-dead-letter" } # Push subscription: inbound-msg → stt-service resource "google_pubsub_subscription" "inbound_to_stt" { name = "inbound-to-stt" topic = google_pubsub_topic.topics["inbound-msg"].name push_config { push_endpoint = "${google_cloud_run_v2_service.stt_service.uri}/transcribe" oidc_token { service_account_email = google_service_account.pubsub_invoker_sa.email } } ack_deadline_seconds = 60 dead_letter_policy { dead_letter_topic = google_pubsub_topic.dead_letter.id max_delivery_attempts = 5 } }

# Alert: STT P95 latency > 6s resource "google_monitoring_alert_policy" "stt_latency" { display_name = "STT P95 latency exceeded" combiner = "OR" conditions { display_name = "STT P95 > 6000ms" condition_threshold { filter = "resource.type=\"cloud_run_revision\" AND metric.type=\"run.googleapis.com/request_latencies\"" comparison = "COMPARISON_GT" threshold_value = 6000 duration = "300s" aggregations { alignment_period = "300s" per_series_aligner = "ALIGN_PERCENTILE_95" } } } notification_channels = [google_monitoring_notification_channel.oncall.name] } # Alert: monthly spend projection > ₹1500 resource "google_billing_budget" "monthly_budget" { billing_account = var.billing_account_id display_name = "AutoHR monthly budget" budget_filter { projects = ["projects/${var.project_id}"] } amount { specified_amount { currency_code = "INR" units = "2000" # hard ceiling } } threshold_rules { threshold_percent = 0.75 } # alert at ₹1500 threshold_rules { threshold_percent = 1.0 } # alert at ₹2000 }