Experts know the process of building a document (the questions to ask, the order of operations, the edge cases), but translating that into a long system prompt often leads to hallucinations or missed steps.
What is Rakenne? Rakenne is a multi-tenant SaaS that lets domain experts define "Guided Workflows" in Markdown. An LLM agent then runs these workflows server-side, conducting a structured dialogue with the user to produce a final, high-fidelity document.
The Tech Stack:
* Agentic Core: Built on the pi coding agent (https://github.com/badlogic/pi-mono) using RPC mode. This allows the agent to maintain state and follow complex logic branches defined in the Markdown files.
* Frontend: Built with Lit web components. I wanted something incredibly lightweight and framework-agnostic so the document "interviews" feel snappy and can eventually be embedded as widgets.
* Multi-tenancy: Designed to isolate agent environments server-side, ensuring that custom expert logic doesn't leak between tenants.
Why this approach? Instead of "Chat with a PDF," it’s "The Logic of an Expert." If you’re a lawyer or a compliance officer, you don’t want a creative partner; you want a system that follows your proven methodology. By using Markdown, we make the "expert logic" version-controllable and easy for non-devs to edit.
I’d love your feedback on:
1. The Agentic UX: Does the "interview" flow feel natural, or is it too rigid? 2. Markdown as Logic: Is Markdown the right "DSL" for this, or should we move toward something like YAML or a custom schema? 3. Latency: We're using RPC for the agent-browser communication—is the response time acceptable for your use case?
Thanks! I'll be around to answer any technical questions.