osmoscraft

SAM 3.1 for segmenting the stickers from the photos

Gemini Embeddings 2 for image and text embeddings

Eigen DB (https://chuanqisun.github.io/eigen-db/) for in-browser vector search

Shout out to https://stickertop.art/ for the collection of photos

Media art project that simulates the open source maintainer experience through an answering machine, using real commit messages from GitHub.

Demo: https://chuanqisun.github.io/eigen-db/

Benchmark: https://chuanqisun.github.io/eigen-db/bench

React hooks can help manage the state and tools for LLM agents. Hooks are good abstraction for dynamically composing data and behavior for both human user and LLM. It also lets human and LLM collaborate on the same UI without communication overhead.

Source code and demos: https://github.com/chuanqisun/react-agent-hooks

I thought it might be easier to piecemeal add LLM functionality if we can reuse the existing structure of the app.

React turns out to be a powerful tool to management the state of anything hierarchical and dynamic.

For the UI, we can show stuff dynamically with JSX; and the JSX would include actions that let user modify the state; the updated the state would trigger a re-rendering of UI to reflect the latest the state.

What if we can apply this paradigm to model Human + AI interacting with a shared piece of state? We can let React and LLM see the same state and use the same tools to change that change.

I built this project to help pay estimated tax. This is useful when I trade stocks that result in unplanned capital gain or income. Before this project, I got myself into tax penalty because I underpaid my estimated tax.

You can probably get the same functionality from Excel, but I think the numbers are more readable when rendered with minimum HTML + CSS.

Sharing it for other hackers who'd like to do their own tax <3

Developer here. I super appreciate the support. Some context:

- I'm a big fan of RSS and there is a thriving community behind. Just check out this list: https://github.com/AboutRSS/ALL-about-RSS

- I'm also a firm believer of web as a platform, a distributed document database, and an open library of knowledge, as opposed to a "compile target" for cryptic JavaScript apps that take control and freedom away from users. Jim Nielsen has a timely critique: https://blog.jim-nielsen.com/2021/web-languages-as-compile-t...

- The bigger picture behind the osmos project is to create an IDE for personal knowledge management. On the surface:

1. I read my rss with osmos::feed.

2. I capture reference links from the feed with osmos::memo.

3. I digest the knowledge and connect them into notes with osmos::note.

- All of them are done with plaintext (some sprinkle of markdown), remote hosted on GitHub, so they are easy to run NLP and ML against. Potentially with GitHub actions, or locally with some bot, with a cloned repo.

- In the long term, I was hoping to create a "positive feedback loop". Use ML to extract patterns from my notes, make connections for me, and recommend interesting reading in the osmos::feed. On the other end, osmos::feed can use NLP to detect how each article in the feed might connect to ideas from osmos::note and make note-taking even easier.

- The parent project (https://osmoscraft.org) is still in super early stage. Would love to let the community give it a spin while I keep iterating.

- Thanks again for the ♥