smart documentation (structify)

/the idea

technical documentation is often fragmented and hard to navigate. the idea was to build a system that actively understands the relationships between different pieces of knowledge, making the docs as smart as the code itself.

/the site

the platform offers a minimal, fast interface for querying documentation. users can ask natural language questions, and the system retrieves not just text, but the contextual relationships surrounding it.

/implementation

the project is built on a modular agentic orchestration pattern, acting as a digital "firm" of specialized ai workers. when a user uploads a system requirements specification (srs) through the frontend, the file is sent to the fastapi backend where a "lead agent" analyzes the document and breaks it down into sub-tasks.

instead of a linear generation process, the lead agent delegates work to a "swarm" of specialized ai agents powered by the google genai sdk. as these agents work in parallel (e.g., parsing entities, generating mermaid diagrams), the backend streams their logs in real-time to the frontend using server-sent events (sse). the frontend seamlessly displays the raw input alongside live agent logs and interactive diagrams within a bento grid ui, providing full visibility into the ai's architectural decision-making process.

/the tech stack

/challenges

the main hurdle was maintaining context across heavily fragmented technical documents. parsing code blocks, maintaining hierarchical structure, and ensuring the llm didn't hallucinate api endpoints required strict chunking strategies.

/thoughts

building structify reinforced my belief that context is everything in ai. an llm is only as good as the data it retrieves. this project was a deep dive into building systems that actually learn the structure of the data they consume.