Production-grade DAG-based AI workflow orchestration platform built with Next.js App Router, strict TypeScript, TailwindCSS, and ShadCN-style component primitives.

• DAG validation with Zod: cycle detection, required-input checks, strict handle type compatibility.
• Deterministic level-based scheduler:
  • same dependency level runs in parallel
  • dependent levels run sequentially
• Strict orchestration/execution split:
  • orchestration in lib/engine/*
  • worker task handlers in lib/tasks/*
• Trigger-style task mapping per node:
  • LLM -> Gemini (@google/generative-ai)
  • Upload Image/Video -> Uploadcare -> Cloudinary persistence
  • Crop / Extract Frame -> FFmpeg worker abstraction
• React Flow workflow studio:
  • collapsible node palette
  • drag-and-drop node creation
  • animated typed edges
  • MiniMap + zoom/pan
  • inline node outputs
  • reactive inputs (connected handles override + disable manual entry)
• Run history panel with node-level logs and statuses for full/partial/single runs.

• Next.js 14 (App Router)
• TypeScript (strict mode)
• React Flow
• TailwindCSS
• ShadCN-style UI components
• MongoDB + Mongoose
• Zod

• app/ -> pages + route handlers
• components/ -> UI and workflow canvas components
• lib/nodes/ -> node contracts, registry, demo graph
• lib/engine/ -> validation, scheduler, run/workflow stores
• lib/tasks/ -> task dispatch and worker handlers
• lib/services/ -> Gemini, Uploadcare/Cloudinary, FFmpeg adapters
• lib/db/ -> Mongo connection and models
• trigger/ -> task exports for worker runtime boundary

1. Install dependencies:

npm install

2. Configure env:

cp .env.example .env.local

3. Start MongoDB locally or provide hosted Mongo URI.

4. Start app:

npm run dev

• MONGODB_URI
• MONGODB_DB_NAME
• GEMINI_API_KEY
• UPLOADCARE_PUBLIC_KEY
• CLOUDINARY_CLOUD_NAME
• CLOUDINARY_API_KEY
• CLOUDINARY_API_SECRET

The app bootstraps a prebuilt Product Marketing Kit Generator workflow with:

• Branch A: image upload -> crop -> LLM description with text prompts
• Branch B: video upload -> frame extraction
• Convergence: final LLM consumes branch outputs for social marketing copy

1. Push repository to GitHub.
2. Import into Vercel.
3. Add environment variables in Vercel dashboard.
4. Deploy.

• Upload and FFmpeg handlers currently expose production integration points and can be routed to dedicated Trigger.dev worker infrastructure.
• Scheduler and worker handlers are intentionally decoupled for horizontal scaling.