Delectable AI Architecture & Technical Evidence
For CTO · CDO · CSO
Reference architecture · v2026.05

Six layers. One platform.
Receipts at every step.

The platform behind the conversation. Composable from foundation to delivery, ratified contracts between layers, per-tenant secret isolation, NIST AI RMF mapped in code (not on a slide), and 10+ tenants live today. Each layer below shows the actual modules, the executive one-liner, and the on-disk evidence.

For the CTO For the Chief Data Officer For the Chief Security Officer For the Head of AI
01
Layer 1 · Foundation Platform-core. Two clouds. Same code.
infrastructure db identity file_management analytics

Config, identity, persistence, files, observability — same code, two clouds (GCP DelectableAI · Azure EIS), per-tenant secret isolation enforced by the env resolver, dual-remote on every module. Configuration is hierarchical: global → customer → environment → module, with secrets stored per-cloud in Secret Manager / Key Vault.

10+ tenant JSONs · infrastructure/customers/ Secret path · via/<tenant>/<env>/<module>/<key> Per-cloud Terraform · deploy/terraform/{gcp,azure}/{dev,stage,prod}/ Audit + provenance · audit_store.py
02
Layer 2 · Integration & Eventing Declarative orchestration. Typed contracts. Event-driven.
workflow communication scheduling calendar safety

Services ship actions.yaml manifests. Workflow ingests them at boot, validates Pydantic schemas, and executes HTTP steps gated on a shared event bus. No tight coupling — composition by contract. The contract was ratified 2026-05-22 and is mirrored by every shared-service module.

Contract · parsing/actions.yaml Multi-cloud secrets · scheduling/secrets_{azure,gcp}.py Safety as a service · safety/routes/{incidents,policies,evaluations}.py Event bus topics declared in each actions.yaml
03
Layer 3 · Knowledge Ingest Six parsers. Four research connectors. Eight social integrations.
parsing sourcing research social

Specialized parsers (Tika · OCR · VLM · video · chem · math) unified by one orchestration surface. Research connectors hit PubMed · bioRxiv · Cochrane · SEC EDGAR · BLS · B Corp · CDP. Social listening from Instagram · TikTok · Google Trends · Digital Shelf. Right tool per format; one typed contract for all of them.

Parser fleet · parsing/parsers/{tika,ocr,vlm,vision-llm,chem,math,video}/ Research · research/connectors/{arxiv,food_research,market_intel,sustainability}.py Viral intent listener · delectable/social/workers/viral_intent_listener.py Evidence quality grading at ingest time
04
Layer 4 · Knowledge Layer Canonical truth. Each module owns one data type.
pim · products cms · content + branding ia · hypergraph governance · frameworks

PIM owns products + food intelligence + embeddings. CMS owns content + branding (per-tenant stylesheets resolved at request time). IA owns the Memgraph hypergraph with provenance per edge. Governance owns frameworks — NIST AI RMF mapped in code, AIRR scoring deterministic, evidence collected automatically.

Hypergraph · ia/knowledge_graph/{graph_models,provenance,measurement_store}.py NIST AI RMF in code · governance/assessment/nist_ai_rmf.py Memgraph deploy · infrastructure/deploy/memgraph/ per cloud × env Tenant branding · cms/branding/<slug>/stylesheet.css
05
Layer 5 · Intelligence Composed score functions. Twelve disciplines. LLM-as-judge.
cart-builder recommender orchestrator propensity scorer 12 grocery domains governance evaluation

Recommenders are composed score functions — not a black box. Twelve grocery sub-disciplines: food science, constraint, householding, pantry intelligence, personalization, semantic bridge, contextualization, consumption, fulfillment, data quality, embeddings, dietary. LLM-as-judge harness grades the agents on persisted runs.

Recommenders · ml/recommenders/{orchestrator,score_functions,agent_tool}.py Propensity · ml/propensity/{session_propensity,signals,train_weights}.py Eval harness · governance/evaluation/runner.py + 4 governance agents Safety redactor in the loop · delectable/ml/safety/
06
Layer 6 · Delivery One API surface. One chrome. Mobile reuses web's agent layer.
api_router console chat-service mobile · @via/mobile partner portal 8 demo portals

FastAPI gateway fans out to module backends. One platform chrome across every console. Expo React Native mobile reuses the same agent layer the web does. Per-tenant branded surfaces ship from one codebase.

Global chrome · workflow/static/platform-chrome.js Mobile agent · delectable/mobile/lib/agent.ts (shared with web) UCP gateway · agent-callable endpoint per tenant 8 tenant portals · demo/public-*
For the skeptical auditor · receipts on disk
Module Registry table (Layer column per service) MODULE-STATUS.md (vctl-generated, reality vs intent) 10+ tenant JSONs in infrastructure/customers/ Memgraph deploy scripts per cloud/env parsing/actions.yaml step-action contract NIST AI RMF mapped in code, not slides Per-tenant Secret Manager / Key Vault paths Identity + entitlements service · audit_store.py session_audit.py in governance · SOC-2-ready foundations Dual-remote git: origin (GCP) + earley (Azure)
Why this matters · the platform story

You're not buying AI. You're buying a contract.

Every grocer has AI vendors knocking. The story is that Delectable's foundation through delivery is one composable platform with ratified contracts between every layer — so when you want to swap in your own analytics warehouse, route through your own identity provider, or bring your own LLM gateway, the contracts hold. That's the difference between a vendor demo and a platform you can build on for ten years.

Where to go next