Gyrus: An Agentic Browser to Increase Intelligence

• 

  logo

• 

  chilling in the hotel lobby @ 3am

• 

  mermaid diagram

Inspiration

AI has totally taken over the world in the past couple of years. People are calling it the “new electricity,” and there’s billions of dollars going into it. Most folks—investors, developers, and everyday users—think it’s the future. But something that’s not talked about enough is how relying on AI too much can mess with how we think. The more we get quick, decent answers from AI, the less we stop to think for ourselves.

We built Gyrus to help with that. Instead of using AI to replace thinking, we use it to support thinking—especially for learning, researching, and creating. And where better to put that than in the one thing everyone uses all the time? A browser.

What It Does

Gyrus is a smart browser that helps you organize your thoughts and learn faster without doing all the thinking for you. It uses something we call crews—basically little AI teams that personalize your experience based on what you’re trying to do.

It figures out your intent by analyzing your past searches and interactions using a memory graph and a fine-tuned model. Based on that, it helps guide you through whatever task you’re working on—research, studying, or just trying to understand something. And don’t worry, when we send info to AI models, we run it through an open-source privacy tool called CHAAP to keep your data safe.

How We Built It

• We built the browser using Electron.js (yeah, we know…).

• Our backend runs on ngrok and connects to tools like:

  • GDELT and NewsAPI for current events
  • Exa for smarter search
  • CrewAI to manage agent teams
  • Langchain for memory and reasoning

• We fine-tuned a Zero-Shot Classifier using the ORCAS-II dataset (about 2 million entries) to understand user intent

• We used Weave to track everything going on between the AI agents

Challenges We Ran Into

• We had a bunch of dependency issues between CrewAI and Weave, so we had to do some package wrangling.
• Exa was super sensitive to how queries were written, which made intent detection tricky.
• Getting our classifier to work on a huge dataset and plug it into everything else wasn’t easy—we had to do a lot of tweaking and data cleanup.
• We are not from Cincinnati so we had to stay in a hotel lobby overnight! Shout out Nikolai at Hilton!

What We’re Proud Of

• Actually getting our classifier to run on a massive dataset and work in real-time
• Getting intent detection + memory graph working together
• Building a functional, private-by-default AI-powered browser in just a few days

What We Learned

• We got way better at working with MCP and LLM agents
• Learned how to fine-tune models for real-world tasks
• And most importantly—how to keep AI helpful without*making people think less

Built With

• arxiv
• classifier
• crewai
• css
• electron
• exa
• gdelt
• html
• langchain
• neo4j
• newsapi
• openai
• python
• roberta
• semanticscholar
• weave