Cerelyze

Turn technical research papers into useable code

125 followers

Turn technical research papers into useable code

125 followers

•

•

Cerelyze allows engineers to quickly reproduce complex algorithms from the latest research papers. We do this by automatically converting methods in the papers into runnable code.

Free

Launch tags:Developer Tools•Artificial Intelligence•Science

Launch Team

Render Sandboxes: a computer for every agentGive every agent its own machine. Starts in under a second

Promoted

SuperCraft

Maker

📌

Thanks for the hunt, Michael! Hi PH! 👋 I am Sarang, Founder of Cerelyze. In my previous life, I was a computer vision engineer building fitness trackers, shopping carts and checkout kiosks. Searching for research papers was the default way to look for novel solutions to complex problems. And yet as I used to search on sites like scholar, arxiv, paperswithcode - it was quite rare to find papers with a github repository that just runs out of the box. Most of the papers just outright did not have any code, and the tiny fraction that had code never ran out of the box. (Some of them required me to downgrade my CUDA version.. like Hello?? 🙄 ) All I wanted was a quick example run of what the paper is trying to do, not in text, but in code, so that I can run it through my own sample data to see if it's even worth investigating. Kinda like HuggingFace, but not just for running ML models, but for running the entire approach of the paper end to end. So I started with dockerizing existing codebases and building a catalog of these "guaranteed to run" images. I showed it to a few folks and got some interest. But I noticed something interesting. Most of the people who showed interest were NOT into computer vision. In fact, their fields of expertise were very different than the usual ML crowd that I used to digitally hang out with. These were people who were trying to reproduce software approaches from papers in fields like Drug Discovery, Biotech, Simulations, Material science, and even art! And the fraction of papers that come with code is even lower in those fields. After talking with them, it was clear that there's a big bottleneck between published research and research that finds its way into commercial products. So after a lot of caffeine, and a couple GPTs later -> Cerelyze was born. It's still in beta so things don't work all the time, but the underlying problem being worked on is to remove that bottleneck. Cerelyze currently is a browser-based environment where users can upload papers, and generate code for specific methods discussed in the paper. Users can also run the generated code directly in the browser. Under the hood is a combination of powerful parsers built for scientific content and LLMs that understand code. It also enables users to ask specific questions about the paper in natural language, making it easy to understand. So feel free to use it, give any feedback, join our discord, and stay tuned! Its completely free and can be accessed at: https://cerelyze.com

Report

3yr ago

SuperCraft

Maker

Thank you! @jackson_huge

Report

3yr ago

Upsolve AI

This is sick! I revisited some of the research papers I used to read/work on when I was in Academia, and looked at the output, from a brief glimpse, it looks decent/legit! Would you support papers outside of arXiv soon? Keen to see if it wrote better code than me for the papers I published. Also, does it support non-Python code output (thinking Matlab, Fortan etc)?

Report

3yr ago

SuperCraft

Maker

Haha thank you! We are constantly working to make the output better. Currently it does not support outside of arXiv, but expect to be able to upload any PDF in the coming weeks. And I personally don't think its going beat the authors themselves in its current state. Currently only python, its just easier to start with that way. But don't see why other languages won't be possible. @kalingwu

Report

3yr ago

CloudFunnels AI

Very interesting... How is the effectiveness in your opinion. Did you create a working app with this?

Report

3yr ago

SuperCraft

Maker

@cyriljeet It really depends on the quality of the paper too. For papers that exhaustively describe the method, its very effective. For papers that are review papers, that do not fully describe the method, it does a poor job because there is no info in the paper. In that case I personally use it to get the psuedocode of the method. For me the pseudocode is a better way to read the paper than reading the text, personally. Overall it tends to be more effective for famous topics like attention mechanism. And yes, I have built functional solutions using this, however in its current state, that needed some manual tweaks to the generated code. hope this answers your question.

Report

3yr ago

Very cool! Regarding code generation, is there something special about research papers like the level of detail that makes it work better with this route as opposed to just arbitrary user-provided descriptions? We're working in a similar space (the latter), but planning and making sense of ambiguities is hard. Looking forward to see how you're able to evolve this.

Report

3yr ago

Zefi.ai

this is amazing , definitely going to test it out! In our startup we spend tons of time reading papers and being on the edge of LLM and ML, this would save us real money. i'll check it out and give you feedback as soon possible! :)

Report

3yr ago

Hi @sarangzambare ! I'm really excited about the potential of Cerelyze. As someone who often finds themselves exploring research papers, the idea of a tool that can generate running code directly sounds amazing. A question - How accurate are the code snippets generated by your system? And how does it handle different coding languages? Thanks for bringing this to the community and looking forward to seeing the project evolve!

Report

3yr ago

SuperCraft

Maker

@jasdeep_chhabra Do see my reply above to a similar question. For some of the more difficult papers, it is able to generate a psuedocode if not a working implementation. It currently only supports python, but if there is enough pull from users, happy to add support for other langs.

Report

3yr ago

@sarangzambare Sorry! I missed that question! Great product btw!

Report

3yr ago

Voxme: AI coach, insights, guidance

How does Cerelyze achieve this seamless conversion, and what inspired you to create a tool that bridges the gap between research and implementation? What a great product!

Report

3yr ago

1 2

Reviews

No reviews yetBe the first to leave a review for Cerelyze