How We Built an AI Generation Engine for a SaaS Creative Suite

Most SaaS platforms add AI as an afterthought — a chatbot in the corner, a "summarize" button nobody clicks. We took a different approach: AI generation is the core value engine of Gigaviz Studio. Every document, chart, image, video storyboard, music composition, and dashboard can be generated from a single prompt.

This post covers the architecture decisions, the code patterns, and the lessons learned from building a unified AI generation system that serves 6 different content types inside a multi-tenant SaaS platform.

The problem: 6 content types, 1 interface

Gigaviz Studio is a creative suite with three modules — Office (documents), Graph (charts, dashboards, images, videos), and Tracks (music). Each module has its own database tables, its own API routes, and its own detail pages.

The challenge: how do you add AI generation to all 6 content types without duplicating logic, breaking the existing CRUD patterns, or making the token economy impossible to manage?

Architecture decision: unified generation endpoint

We rejected the obvious approach of adding generation sub-routes to each CRUD API. Instead, we built a single unified generation endpoint that handles all 6 types through a discriminated union schema.

One schema. One validation path. Six content types. The discriminated union gives us type safety at the boundary — TypeScript knows exactly which fields exist for each type after validation.

The generation pipeline

Every generation request follows the same 5-step pipeline:

1. Validate → The discriminated union schema parses the request body 2. Authorize → Check user session, workspace membership, and entitlement (`graph`, `office`, or `tracks`) 3. Deduct tokens → Call an atomic token deduction function before doing any AI work (fail fast on insufficient balance) 4. Generate → Call the appropriate AI function (GPT-4o-mini or DALL-E 3) 5. → Update the entity's status to `completed` and write generated data to the database