— Surface 01 · Architecture · Live · From gcloud asset list

Not a slide.
The actual graph.

Drawn from this morning's gcloud asset list --project=delectable-cloud-stage — 1,780 assets, 70 distinct Google services in production. The production graph for the grocer's chief architect — every service, every dependency, every dataplex, traceable to the live project.

Project: delectable-cloud-stage
Project number: 1057394044627
Org: 707792343664 · delectableai.com
Inventory pulled: May 25, 2026 22:50 PDT
— Inventory snapshot · live data

Seven dimensions of GCP presence.

1,780
Total assets
70+
Distinct services
22
Cloud Run services
97
BigQuery tables
5
BigQuery datasets
526
Cloud Run revisions
14
GCS buckets
— The GCP-icon view · generated from Delectable's own cloud-architecture agent

Familiar icons. Production-real graph.

Same architecture as the layered breakdown below, but rendered in the official Google Cloud iconography the GCP team uses internally. Generated by our infrastructure/cloud_architecture module — the same module a Delectable engineer (or agent) uses to specify and emit Terraform for every new grocer tenant. The diagrams are code, not Figma exports.

i. Five-layer overview · every Cloud Run module + every Google service it touches.
Delectable AI production GCP architecture
Click to enlarge · official Google Cloud icons via mingrammer/diagrams
ii. Agentic shopper request · 490ms end-to-end through the same stack.
Delectable agentic shopper request flow through GCP
Numbered: 1) ground in HyperGraphs · 2) reason with Gemini · 3) retrieve via Vertex Retail Search · 4) price + stock from Cloud SQL · response in 490ms.
— The production graph · Five layers, top-down

Edge to data. All Google.

This is the graph an agentic shopper request traverses today — measured against the live asset inventory, not redrawn for the meeting. Read top-to-bottom.

GCP infrastructure
Gemini / Vertex AI
Vertex Retail / Discovery Engine
Firebase
Delectable IP layer
01 Edge & identity 7 services · 1 LB · 2 HTTPS certs · 1 WIP
Firebase App Hosting
delectable-prototypes
delectable-partner
delectable-dai · +21 sites
3 services · 1 project
Firebase Auth + Identity Toolkit
OAuth IDP · OpenID configs
delectableai.com SSO
1 default + 1 OAuth IDP
Cloud Load Balancing
2× HTTPS target proxies
4× URL maps · 2× SSL certs
Global · multi-region
↓ request flows ↓
02 Capability modules 17 capability surfaces · every module is a composable platform leg · each consumes a distinct GCP service mix
i. Tenant-domain — what the grocer's shopper, merchandiser, marketer, advertiser sees.
Agentic Chat
The Perfect Cart · grocery agent · 490ms baskets
Cloud Run · Gemini Pro · Memorystore · pgvector
Search & Discovery
Grocer.com semantic search · agentic browse
Vertex Retail Search · Discovery Engine · Vector Search
Ads / RMN
Retail-media network · supply & demand · ROAS closed-loop
Cloud Run · BigQuery · Google Ad Manager · Google Ads API
Social Commerce
TikTok-style shoppable feed · creator commerce
Cloud Run · GCS · Firestore · YouTube Data API
Console & Partner Portal
Admin · merchandiser · marketer · partner · FSR portals
Firebase Hosting · Vite/React · Firebase Auth
ii. Knowledge-domain — modules that own a canonical type of structured data.
PIM
Product information · 8 → 42 attribute lift · Merchant Center conversational attrs
Cloud Run · BigQuery · Vertex AI · GCS
IA · Information Architecture
27 ontologies · dietary · allergen · cultural · mission
Cloud Run · BigQuery · Memgraph · Delectable IP
CMS
Multi-tenant content · branding · recipe + editorial
Cloud Run · Firestore · GCS · Cloud SQL
Sourcing
Vendor catalog ingest · Salsify · Akeneo · SAP · VTEX · plus Azure / AWS / Databricks bridges   ↗ flows
Cloud Run · Pub/Sub · GCS · BigQuery
Governance
AI governance · SOC 2 · model lineage · audit
Cloud Run · BigQuery · Dataplex · Cloud Logging
iii. Shared services — composable capability legs · step-action manifest pluggable.
Workflow
Step orchestration · retry · gating · multi-step agents
Cloud Run · Cloud Tasks · Pub/Sub · Firestore
Orchestration
LangGraph agent runtime · tool routing · conversation state
Cloud Run · Memorystore · Gemini Pro
Communication
Outbound notifications · transactional email · push
Cloud Run · Gmail API · Pub/Sub · SendGrid
Scheduling + Calendar
Booking flows · OAuth-connected calendars · delivery windows
Cloud Run · Google Calendar API · Cloud SQL
Safety
PII redaction · content moderation · prompt guardrails
Cloud Run · Vertex AI · Data Loss Prevention API
iv. Platform-core & infrastructure — assumed by everything else.
Identity
Multi-tenant auth · entitlements · RBAC · SSO
Cloud Run · Firebase Auth · Identity Platform · IAM
Analytics
Shopper analytics · basket analytics · cohort reporting
BigQuery · Looker · Dataform · Data Studio
API Gateway
Public API surface · UCP dispatch · rate limit · key mgmt
Cloud API Gateway · Cloud Endpoints · Apigee (planned)
Event Bus
Inter-module events · sourcing · workflow · governance
Cloud Pub/Sub · Kafka (Confluent on GKE) · Eventarc
Testing / Eval
LLM-as-judge eval harness · contract tests · cross-tenant smoke
Cloud Run Jobs · Vertex AI Eval · BigQuery
▸ The yellow-bordered modules consume Google APIs that bill outside the Cloud Sales P&L — Ad Manager / Google Ads (Google Ads team), Gmail + Calendar + YouTube + Maps (Workspace + Maps team). Same grocer deal lights up multiple Google P&Ls. See Broader Google footprint →
↓ reasoning + retrieval ↓
03 AI · reasoning · retrieval Vertex AI · Gemini Pro · Vector Search · Discovery Engine · Retail Catalog
Gemini Pro
Reasoning model for the 20% of grocery queries that require deep inference. Called after Delectable grounds the prompt.
Vertex AI Reasoning Engine
Managed agent runtime · tool use · multi-turn orchestration · session memory.
Vertex Retail Search
Live on the grocer's catalog. Semantic relevance for grocer.com search — outperforms Algolia / Constructor on Delectable-enriched corpora.
Vertex Discovery Engine
Multi-region collections for cross-tenant, cross-locale agentic search. The retrieval substrate UCP requests fan out to.
Vertex Vector Search
Embedding lookups for find_products, find_alternatives, recipe matching. 3 lookups per agentic basket build.
Food HyperGraphTM
142k ingredients · 9 dietary vectors · 27 cultural axes. Queried before Gemini — eliminates SKU hallucination.
Delectable IP · patent pending
Shopper HyperGraphTM
847 signals per household. Joined to BigQuery as a federated source. 94% prediction accuracy on recurring baskets.
Delectable IP · patent pending
Memgraph (KG)
Property graph backing the HyperGraphs. Products · ingredients · recipes · dietary flags · occasions — all connected.
Runs on GKE · GCS backups
IA ontology layer
27 cross-cutting taxonomies. Dietary · allergen · cultural · mission · life-stage · occasion. Reviewed by food scientists per release.
Delectable IP
CCAI & Dialogflow ROADMAP
Planned · 2026 H2. Customer-service intents — order status, returns, product Q&A, dietary advisory. CCAI Insights for call analytics. Dialogflow CX for IVR + chat fallback. Same grocer deal, more Google services consumed.
Agent Builder & Gemini Enterprise CX 2026
Drop-in integration target. Delectable's HyperGraph layer makes Gemini Enterprise CX actually accurate for grocery — closing the gap that currently blocks Tier-2 grocer adoption.
↓ data + storage ↓
04 Data & storage 5 BQ datasets · 97 tables · 14 buckets · 2 SQL · 1 Redis · 1 Firestore
BigQuery · food_intelligence
Food HyperGraph store
Ingredient × nutrition
BigQuery · retail_analytics
Loyalty + transaction joins
Shopper signals · 847/household
BigQuery · enrichment_warehouse
PIM input/output
Catalog lineage
BigQuery · retail_media
Ads impressions + ROAS
Closed-loop attribution
Cloud SQL (Postgres + pgvector)
delectable-cloud-stage
chat-service-staging-db
+ pgvector embeddings
Memorystore for Redis
chat-service-staging-redis
Session + cache
Firestore (NoSQL)
Realtime shopper state
1 default DB
GCS · 14 buckets
delectable-food-data
delectable-cloud-stage-data
via-external-datasets · +11
Plus Dataplex · 11 entry groups + Dataform · 1 repository (governed data plane).
↓ devops · iac · governance · security ↓
05 DevOps · Infrastructure-as-Code · Security Terraform-deployable in 15 min · GitHub-native CI · Cloud Build · Secret Manager · IAM
i. Infrastructure-as-Code — every grocer tenant stands up the same way.
Terraform · customer-tenant
One bundle, one ./apply.sh, 15 minutes. Provisions Cloud Run, BQ, Cloud SQL, Secret Manager, IAM, AR — all in the grocer's GCP project.
commerce/deploy/customer-tenant/
Cloud Build + Triggers
Container build farm. GitHub-connected triggers per service. Cloud Native Buildpacks for Python/Node services.
Cloud Deploy
Progressive rollouts dev → stage → prod. Canary + manual-gate stages. Per-tenant deployment targets.
Artifact Registry
8 Docker repos · 550 versioned images. Vulnerability scanning on every push. Signed by Binary Authorization.
ii. CI/CD & GitOps — code-to-prod in under 12 minutes.
GitHub Actions
Per-module matrix workflow · lint, type-check, test on PR · matrixed cross-tenant deploys on merge.
Workload Identity Federation
GitHub Actions → GCP without long-lived service-account keys. The auth pattern Google publishes as best practice.
Dataform
SQL-as-code for the BigQuery analytics layer. Version-controlled transforms, dependency DAG, scheduled runs.
Cloud Scheduler + Tasks
Nightly enrichment refreshes · daily catalog diff · weekly Shopper-HyperGraph regeneration.
iii. Security & observability — every grocer's CISO sees this slide.
Secret Manager
Per-tenant scoping · CMEK encryption · automatic rotation. No secrets in code, no secrets in CI YAML.
IAM · Org Policies · WIP
18 service accounts · 1 Workload Identity Pool · 3 enforced org policies. Least-privilege by default.
Cloud Logging + Monitoring
Structured logs · BigQuery export · 2 log sinks. Cloud Monitoring uptime checks + alerting policies.
VPC · Subnets · Firewalls · LB
3 networks · 45 subnets · 8 firewall rules · 1 router. Cloud Run egress via VPC Connector for private Cloud SQL reach.
Every box above maps to a real asset in the inventory dump (gcp-inventory.json). No marketing logos. No "future state."
— Agentic Infrastructure-as-Code · the meta-story

Delectable uses agents to define and ship
its own GCP infrastructure.

The cloud-architecture diagrams above weren't drawn in Figma. They were generated by Delectable's infrastructure/cloud_architecture module — Python code that an engineer (or an agent prompt) emits. Same module emits the Terraform that provisions a grocer's GCP project. Same module emits the asset-inventory validation harness. If you want to see what agents can do for the grocer's IT team, look at what they already do for ours.

Delectable agentic IaC and CI/CD pipeline
Same diagram generator. Three persistent outputs: the architecture PNG (above), the Terraform plan, the customer-tenant deploy bundle.
01
— What the agent does

Specifies infra in code.

An engineer prompts "stand up a Tier-2 grocer tenant in us-east1 with 8 Cloud Run services and a 4-vCPU Cloud SQL." The cloud-architecture module emits a Python spec → Terraform plan → graphviz diagram.

02
— Why Google wins

Vertex generates the spec.

The next version of this loop (2026 H2) uses Gemini Pro to generate the Python spec from natural language. Every grocer tenant a Delectable agent stands up = a Gemini API call billing to the grocer's tenant. Agentic IaC is itself a consumption motion.

03
— What the grocer gets

Pilot in 15 minutes.

One ./apply.sh command provisions Cloud Run, BigQuery, Cloud SQL, Secret Manager, IAM bindings, Artifact Registry — all in the customer's GCP project. Reproducible. Versioned. Idempotent. Auditable.

▸ For the GCP Solution Architect in the room: The module is at ~/code/infrastructure/src/infrastructure/cloud_architecture/. Multi-cloud abstractions in base.py; GCP services in gcp.py; Terraform emitters in terraform_generator.py. Interactive console at generate_cloud_architecture.py. Apache 2-licensed within the Delectable repo. Open for joint development if the partnership ships.
— Two views of the same stack

21 logos for the boardroom.
70 distinct services for the architect.

Both views matter — the logo grid earns the boardroom in 60 seconds; the production graph wins the working session with the grocer's CTO. This page is the architect view, including the seven Vertex AI Retail Search collections, the three Discovery Engine collections, the Reasoning Engine in us-west1, the Dataplex governance plane. The conversation that gets to a signed pilot.

— Boardroom view · 21 Google services

The logo grid.

For the boardroom and the FSR manager. "We are a Google company." Firebase. Cloud Run. Vertex. BigQuery. Gemini Pro. Five categories, 21 logos, one stack. Closes the credibility loop in 60 seconds.

— Architect view · 70 distinct asset types · 1,780 assets

The production graph.

For the grocer's chief architect. Dataplex entry groups + Dataform repos prove governed-data discipline. Workload Identity Pool proves Cloud-Run-to-BigQuery security model. Retail Catalog asset proves we're already production-running the same product their team is being asked to consider.

— Why each layer matters to Google's P&L

Every Tier-1 grocer Delectable lands brings the full graph with it.

The numbers below scale linearly with the catalog size and the shopper traffic. A Tier-2 grocer (~70k SKUs, ~250k weekly active shoppers) brings ~$28k/mo of consumption on day one. A Tier-1 grocer (Kroger, Albertsons) brings ~10× that.

Layer 2 · Compute
22 services
Cloud Run autoscaling — paid per millisecond per shopper request. Grocer-sized fleet doubles every Tier-1 grocer landed.
Layer 3 · AI · Net Vertex consumption
$1.6M / yr
Tier-1 grocer annualized Vertex AI + Cloud Run + BigQuery spend that exists because the 14× per-call efficiency lets the agentic basket actually ship. Without Delectable, that workload is at zero.
Layer 4 · Data
2.94M
Data points per grocer catalog after enrichment. 5.25× grounding lift on Gemini calls — Vertex stays sticky.
— For your CTO

The full asset inventory, as JSON.

gcloud asset list output, structured by service category. Reviewable by your team. Use it to ground the partnership architecture conversation.

↓ gcp-inventory.json ↗ Open in GCP Console