Gijs
357 posts
@datagobes
Senior cloud data & AI engineer. Building https://t.co/SGOypsko6x in public — AI-native tools, MDX blogs, and things that probably shouldn't work but do.
Andrej is right. Processing pdf's is hard but lots of knowledge is captured in them. I am focusing on using AI in a corporate setting and document intelligence is a challenge. I ran into this problem when I first started designing an agent that could design agentic systems. I wanted to put all the big papers in a knowledge base and it didn't work out. Until I created Felix, my document intelligence project. Felix decomposes any business document into typed elements, stores them in postgres, and can reconstruct the full document from it. Now that Andrej dropped his llm-wiki idea, I've added an mcp on top so I can use Felix for llm-wiki generation and I will post my wiki from the Mythos document on GitHub later on. (Claude happened to go down the moment I was generating it) See the screenshots for a UI view on the Felix API, which is showing the document Andrej is mentioning here, decomposed and served from postgres. I am not going to tell you to reply FELIX and follow me to get some nonsense prompt and a 199 offer for a crappy workshop. But I am not going to stop you from letting me know you think this is interesting either. If this gets some engagement I will package it and put it on GitHub.
@kepano I just tried it this morning on the 245-page Mythos pdf and it failed badly and the outputs were all mangled. Converting pdfs is really hard, I think it has to probably be a Skill not a program, for a SOTA LLM for it to work properly.
SOMEONE built an AI agent that sells pool installations on autopilot 10 "boring" cash-flowing startup ideas YOU can build on autopilot using the OpenClaw/Hermes etc: 1. find commercial buildings with flat roofs in sunny states and calculate their solar savings, render the install, mail the building owner a custom ROI report. become the broker between building owners and solar installers, take a cut of every deal or charge $$ 2. find shopify stores doing $1M+/yr with no international shipping and build them a localized storefront for their top non-US traffic countries, pitch a rev share to unlock revenue they're leaving on the table 3. find businesses paying for 10+ SaaS tools via public job postings and tech stack data and build a custom "consolidation audit" showing how to cut 40% of their software spend, sell the migration as a service 4. find commercial properties with high water bills using public utility data and render a xeriscaping or rainwater capture plan with projected savings, sell to property management companies at scale 5. find ecom brands running meta ads to products with 1-2 star reviews and build a better version of their top SKU with a manufacturer, launch against them with their own keyword data 6. find small banks and credit unions with websites from 2012 and render a modern site + mobile app with their branding, pitch it as a turnkey digital transformation. they have budget but no one's calling on them 7. find warehouses and industrial spaces near EV corridors with no charging infrastructure and model the revenue from installing chargers, pitch landlords a lease + install package 8. find franchisees posting complaints in public forums about their franchisor's tech and build a shadow operating system (POS, scheduling, inventory) that plugs into their existing franchise, sell directly to franchisees 9. find medical practices billing under specific CPT codes with low reimbursement rates and build an AI billing optimization engine that reclassifies and appeals claims, take a % of recovered revenue 10. find DTC brands with 100k+ instagram followers but no subscription offering and model their repeat purchase data from reviews, build a subscription flow with retention math, pitch it as a done-for-you program with rev share the framework is basically this use AI agents to surface a gap in public data, build the solution before anyone asks for it, and show up with the math already done. you use info arbitrage + OpenClaw is the ultimate wedge i dont know how long this lasts but i do know there's a ton of ways to make $$ using this and i won't hold back i'll be sharing more ideas here, @startupideaspod and @ideabrowser today is a beautiful day to be building
