The Project
Blockfrost is a leading Cardano API provider that hosts infrastructure so that developers can build on the blockchain without running their own nodes. It is widely adopted by wallets and dApps across the ecosystem. The Blockfrost Platform is an open-source platform that lets independent Cardano node operators run their own Blockfrost-compatible API nodes, forming a decentralized backend for serving the Blockfrost API. During my time at Input Output Global, I was assigned for two to three months to assess and improve Blockfrost Platform’s documentation. The team had not had a technical writer; documentation had been a side project of the developer team.The Challenge
I inherited a blank FAQ page. There were no earlier versions of a FAQ page to draw from. It wasn’t about updating the FAQ, it was about creating a completely new one. I discovered a gold mine of dense technical conversations happening in the team’s Discord. Developers and early adopters were addressing questions and points of confusion. None of it was being captured anywhere that a future developer could easily find it. The raw material existed, just in a format no typical documentation workflow knows how to ingest. With months of backlog across multiple channels, and having limited time to create the FAQ, combing through all of the Discord history by hand wasn’t feeling like the best method.My Approach
So I built a tool. ChatDoc-InsightMiner-PromptLab is a Python toolkit that ingests chat logs into a vector database, runs analytical prompts across three LLMs at once (OpenAI, Anthropic, Gemini), and cross-validates the outputs. If all three models independently flag the same pattern, it’s likely real. If only one does, it may be an artifact. I fed it months of Discord history and asked it to find the recurring themes, questions and points of confusion. After review and editing, we had a rich, categorized FAQ based on problems that real users were asking about. Result: platform.blockfrost.io/faqMulti-Model Validation
Ran findings through three LLMs. Kept only what all three flagged independently.
Pattern Identification
Found recurring questions and confusion points from months of conversation.
Prioritized Output
Ranked FAQ topics by actual user friction, not assumptions about what developers might want to know.
Human-in-the-Loop
The toolkit wrote first drafts. I reviewed and checked every answer against the platform, then got developers to review before publishing.
Information Architecture Improvements
I also restructured the broader platform docs: added conceptual overviews, clarified the developer pages, fixed the navigation. Result: platform.blockfrost.ioTech Stack
Python · Vector Databases · OpenAI API · Anthropic API · Gemini API · Multi-LLM Orchestration · Markdown