Marginalia

Inspiration

Reseach is time consuming. LLMs and agents can help, but sometimes can make it so difficult that it's just them doing work, and humans trying to understand it after (like a certain Navier Stokes problem...). Without the help of AI though, sometimes it just takes way longer than it should. Marginalia helps you speed up the process, while you maintain understanding. I was talking to someone from Jane Street at Hack the North who said they found AI helpful to try out things they didn't have time to, so they could dedicate their time to others, but after the AI compelted the tedious work, they were able to look at the results and continue from there.

What it does

1. Literature Search: Given a topic, it searches academic databases like arXiv to find related papers.
2. Branching: It analyzes the literature and proposes potential research directions.
3. Human Moderation: Review the branches and literature search at each step and take out or add related papers.
4. Autonomous Execution: Once approved, Marginalia dives deeper, writing and running simulation code or compiling targeted findings, and returns a cohesive research report.

How we built it

• Gemini: Provides the AI reasoning: reads retrieved paper excerpts, proposes research branches, generates Python code for your approved experiment, and helps interpret conference information.
• Vultr: Runs the website, Python backend, research workflow, and approved experiment code.
• Tiger Data: Stores investigations, paper metadata, branch plans, results, manuscript drafts, and payment records in PostgreSQL.
• Solana: Handles blockchain payments for conference registration after acceptance. The app prepares an organizer-issued SOL payment, your Phantom wallet approves and signs it, and the backend verifies the finalized transaction.

Challenges we ran into

Getting the agent to actually do the code writing part and finding one that wouldn't take forever to actually generate the code and timeout.

Accomplishments that we're proud of

Seeing the idea of something that could actually be useful start working!

What we learned

The idea was ambitious, but breaking it down into steps helps to actually see what to create.

What's next for Marginalia

Optimizing the branching, and ensuring literature search is thorough.

Built With

• css
• gemini
• html
• javascript
• python
• solana
• tigerdata
• vultr