Zeke

// GitHub

// Problem

After spending 44 days making daily YouTube videos on my second channel, I realized that the AI and marketing content landscape is drowning in fluff. There are thousands of podcasts, papers, videos, and blog posts published weekly—but most of it is recycled hype. I didn't want to spend 8 hours listening to Lex Friedman podcasts or sift through 10 blogs just to find the two minutes that actually matter. The real signal was buried under mountains of noise, and there was no efficient way to separate genuine breakthroughs from clickbait.

// Solution

A 'Bloomberg meets Canva' style workspace where you could see hundreds of thousands of pieces of content boiled down to what actually matters. The core innovation was a hype scoring algorithm that rated each piece of content on relevance and substance—40% keyword match, 30% highlight type (breaking changes outrank generic quotes), 20% source authority (Anthropic announcements beat random blogs), and 10% freshness. Every insight came with timestamps, citations, and receipts so you could verify claims in seconds instead of hours.

// What I Built

A full-stack research platform built on Midday's open-source finance architecture, repurposed for content intelligence. The system ingests content from YouTube, arXiv, RSS feeds, and blogs, then runs it through a multi-stage pipeline: an Engine layer fetches and normalizes content, a Jobs layer (Trigger.dev) orchestrates AI extraction, and a Dashboard surfaces prioritized insights with jump-links to source timestamps. I built speaker-aware outlines for podcasts, novelty detection for papers, and 'Why it matters' briefs that connected findings to user-defined goals. The scoring system automatically surfaced breaking API changes (0.95+ score) above generic quotes (0.4 score).

// Technologies

// Lessons Learned

• 01Building content intelligence is easier than building distribution. The product worked—I could genuinely find signal faster—but I realized I had no desire to market a SaaS in the crowded 'AI productivity' space. The best side projects solve your own problem; the best businesses solve problems you want to keep solving for others.
• 02Forking existing architecture (Midday) was the right call. Their multi-tenant patterns, team permissions, and component library saved weeks of boilerplate. The cognitive dissonance of seeing 'invoice' variable names in a research app was worth the velocity.
• 03Hype scoring is genuinely useful but hard to explain. Users immediately understood 'this podcast has 3 highlights worth watching' but struggled with why one source scored 0.85 and another scored 0.72. Transparency in AI systems often creates more confusion, not less.
• 04Two weeks is enough time to validate whether you want to ship something. The codebase is solid and the architecture scales—but I learned that technical completion and market conviction are completely different things.