Screensharing Kevin Rose's AI Workflow/New App

What It Covers

Kevin Rose demonstrates his complete AI workflow for building a TechMeme-style news aggregator called Nylon in approximately one week. He screen-shares the technical architecture, clustering algorithms, vector embeddings, and editorial scoring system while explaining how solo builders can now create production-quality products using AI coding tools without traditional engineering skills.

Key Questions Answered

• •Multi-source content enrichment pipeline: Rose built a system that pulls from 63 RSS feeds, then enriches each article through three methods—Firecrawl for scraping, iFramely for metadata, and Gemini AI as fallback—selecting the highest quality result. This redundancy ensures 99.8% of stories get properly processed even when individual sources fail or block crawlers, costing under $100 monthly for thousands of daily tasks.
• •Vector embeddings for nuanced clustering: Traditional keyword search cannot distinguish between "Apple sues Google" versus "Google sues Apple," but vector embeddings create mathematical representations of content meaning. Rose uses OpenAI's large embedding model with clustering algorithms to group related stories across sources, then expands clusters by hitting Brave or Tavily search APIs to discover coverage beyond his RSS ecosystem when three-plus sources report the same story.
• •Gravity engine editorial scoring: Rose created an AI judge that scores stories across multiple dimensions—industry impact (90% for NVIDIA's $2 billion CoreWeave investment), consumer impact (10%), actionability (30%), novelty detection, and PR fluff risk. The system identifies paid sponsorships by detecting similar vector distances between articles published within one hour, revealing coordinated but undisclosed promotional content across multiple outlets.
• •Trigger.dev for durable task orchestration: Rather than unreliable cron jobs or edge functions with timeouts, Rose uses Trigger.dev to create TypeScript micro-instances that automatically retry failed tasks—critical when AI models timeout or RSS feeds get blocked. Each task shows complete execution chains with millisecond-level compute costs, enabling local development while production data continues enriching in the background database.
• •Personal software era philosophy: Rose argues the future is building hyper-personalized tools for yourself first, potentially serving 500-1000 similar users rather than millions. He advocates shipping 100% of ideas as messy prototypes, then cutting 90% of features to find the 10% that matters. The hardest problem is not scaling buggy code but finding what people actually want—vibe-coded MVPs solve product-market fit before hiring engineers.