Architecture

Tailormade is a single Next.js 15 app — no separate backend service. All studios, APIs, auth, and the pipeline engine live inside next-app/.

System Overview

Directory Layout

next-app/
  app/
    api/            API route handlers (auth, promotion, animation, pipeline)
    auth/           Custom JWT auth endpoints (token, refresh, logout, me)
    [country]/      Studio pages (uk, de)
      promotion/    Promotion wizard + agentic
      animation/    Video wizard + agentic
      pipeline/     Image Pipeline node-graph canvas
    login/          Login page (password + Entra SSO button)
  lib/
    auth*.ts        JWT creation, verification, token store
    auth-entra.ts   PKCE flow, Microsoft token exchange, JWKS verification
    services/
      geminiService.ts  Vertex AI REST client (text + image)
    animation/      Veo video generation
    pipeline/
      workflow/     Graph executor, SSE progress, cache, storage
      nodes/        Gemini image-gen node, expand node, IO nodes
  src/
    modules/        React feature modules (auth, promotion-studio, video-studio, pipeline)
  prisma/
    schema.prisma   Database schema (User, Workflow, Run, NodeCache, …)

Authentication

Two parallel auth paths share the same JWT/cookie mechanism:

• Access tokens: HS256 JWT, 30 min TTL.
• Refresh tokens: HS256 JWT, 7 days, stored in httpOnly cookie.
• Token blacklist: invalidated tokens stored in PostgreSQL via auth-store.ts.
• New Entra users are auto-provisioned if ALLOW_ENTRA_USER_PROVISIONING=true and their domain is in the allowlist.

Studios

Promotion Studio (/[country]/promotion)

A 7-step wizard backed by Redux Toolkit and a Konva canvas:

1. Country / language selection
2. Product details
3. Background image generation (Gemini image model)
4. Layout composition (Konva)
5. Text generation (Gemini text model)
6. Social copy generation
7. Export

Agentic mode (/promotion/agentic) streams NDJSON from /api/promotion-studio/agentic.

Video Studio (/[country]/animation)

A 6-step wizard:

1. Campaign brief
2. Scene planning (Gemini text)
3. Prompt optimisation
4. Image descriptions
5. Video generation (Veo — Vertex AI us-central1)
6. Review + export

Agentic mode (/animation/agentic) uses a client-side task queue over SSE.

Image Pipeline (/[country]/pipeline)

A React Flow node-graph canvas. Users wire together nodes and execute the graph:

• Engine: lib/pipeline/workflow/ — topological sort, per-node input/signature cache, SSE progress.
• Persistence: Workflow (graph definition) + Run (execution record) + NodeCache in PostgreSQL.
• SSE: client consumes /api/pipeline/workflows/[id]/stream via fetch reader (not EventSource) so the Bearer token flows.

AI Model Routing

Two separate SDK paths reach Vertex AI:

Path	SDK	Used by
lib/services/geminiService.ts	Hand-rolled Vertex REST + google-auth-library	Promotion, Video, Image studios
lib/pipeline/workflow/genai.ts	@google/genai 1.x	Image Pipeline nodes

Both respect the same env vars. GOOGLE_GENAI_USE_VERTEXAI=false switches both to API-key mode.

Database Schema

Table	Purpose
User	Registered users — email, hashed password, isActive flag
Workflow	Saved Image Pipeline graphs (JSON nodes + edges, userId owner)
Run	Execution records — status, outputs, error, graph snapshot
NodeCache	Input-signature → output cache keyed by node type + inputs hash

Request Flow — Image Pipeline Run