The Brief
A second-time founder - whose first business, a PR agency, had succeeded - wanted to build a SaaS product. His background was in PR, SEO, and affiliate marketing. He connected with M87 through a mutual contact and brought Daniel on as CTO.
The original idea was a Wix-inspired widget builder for affiliate network websites: a SaaS where site owners could manage offers across multiple sites, with a custom widget builder for styling and deploying offer displays. The team built it, but the product struggled to gain traction - and then ChatGPT launched.
The timing changed everything. With AI tools suddenly exploding, the team pivoted to what would become Promptchainer: a visual workflow automation platform for composing and executing AI prompt chains. The original widget builder's drag-and-drop UX experience directly informed the new product's visual editor - the pivot wasn't starting over, it was redirecting the technical foundation toward a much larger market.
What We Built
Promptchainer was a no-code platform for building AI workflows visually. Users designed prompt chains by dragging nodes onto a canvas, connecting them, and running complex multi-step AI workflows - no code required.
The Visual Editor
Built on ReactFlow, the editor supported seven node types: variables (input parameters), actions (LLM calls with configurable model, temperature, and token limits), image generation (DALL-E), code execution (custom JavaScript), context nodes (RAG/vector search), conditional logic (branching), and output collection. Users could wire any combination into sophisticated AI pipelines.
The Execution Engine
Under the hood, a custom graph traversal engine resolved node dependencies, interpolated variables across the chain, cached intermediate results, and propagated outputs step by step. This wasn't a toy - it handled complex branching workflows with real error handling and timeout management.
The Architecture
The codebase was a Yarn monorepo with 13 packages, orchestrated by Turbo:
- Web client (Next.js 13) - the visual editor, auth, subscription management, and a published workflow marketplace
- API server (Express) - REST endpoints, a JobManager polling system, and flow compilation services
- Worker pool (Express + BullMQ) - horizontally scalable workers consuming model-specific Redis queues for async execution
- Shared packages - a graph execution engine, model configurations, Zod schemas, a credit calculation system, S3 integration, Prisma database layer, and shared UI components
The full codebase: 721 TypeScript files, ~41,000 lines of code, 14 database models.
Key Technical Features
- Multi-tenant RAG - users uploaded documents, embedded them with OpenAI's embedding models, and stored vectors in Pinecone for knowledge-grounded prompting
- Chain compilation - GPT-4 could transform multi-step chains into optimized single prompts
- Token-level credit metering - tracked prompt and completion tokens per call for subscription billing via LemonSqueezy
- BYOK (Bring Your Own Key) - users could plug in their own OpenAI API keys
- Publish & share - a marketplace where users could publish, version, fork, and like workflows
- Bot framework - Telegram integration for triggering workflows from chat
- API keys - programmatic access for developers who wanted to call workflows from their own apps
The Result
Promptchainer ramped quickly to $2k MRR, driven by visibility on AI tool aggregators and directories - the early movers in cataloging the explosion of AI products in 2023. The platform attracted users who wanted to build multi-step AI workflows without writing code, from content generation pipelines to RAG-powered research assistants.
This was one of the earliest visual AI workflow builders on the market - before tools like Langflow, Flowise, or n8n's AI nodes existed. The architecture, the speed of execution from pivot to paying users, and the bet on AI workflow automation as a category all proved right.