Your vision AI agent to build VisionAI applications
Atlas is an open-source Vision AI Agent built into Picsellia, allowing you to interact with visual data using natural language to explore and improve your Image datasets. No code, no hassle. Simplify your computer vision workflows and let Atlas work for you.
Honestly, we're so excited (and maybe a little nervous!) to finally share Atlas with you all today.
As AI engineers ourselves, we kept hitting the same wall: general AI models are impressive, but quite overkill and shitty for custom VisionAI use-cases. Supervised fine-tuning is still king.
But in the same time LLMs are becoming so good, so we though what if? What if LLMs could help us with data-centric AI?
Humans, paired with AI agents, can curate Datasets, evaluate model trainings, and monitor models waaaay faster.
So, we built Atlas – an open-source AI agent designed to tackle this head-on.
Think of it as the Data-centric expert co-pilot we wished for. It helps you:
Analyze datasets (finding labeling issues, imbalances, etc. – the stuff that trips up models!)
Suggest the right model architectures for your specific task
Train, evaluate, and iterate much faster
Here's why we think it's different, especially for this community:
Speed & Precision: Seriously, our goal is to help you go from data to robust custom models way quicker.
It's Open Source: Because we believe in transparency. No black boxes – inspect it, fork it, help us make it better! We want this to be a community effort.
LLMs never sees your Data: This was non-negotiable for us. Atlas uses abstract representations and never sends your actual images externally. Your competitive advantage stays safe.
Atlas can integrate with platforms like Picsellia (where we initially built it!) or run standalone (coming soon). Honestly, we built Atlas for people like us: ML engineers, developers, researchers tired of reinventing the wheel for every new vision project.
We've poured our hearts (and countless hours!) into making powerful VisionAI development more accessible and transparent for everyone.
We'd be thrilled if you check out the repo, give Atlas a spin, and let us know what you think! Seriously, all feedback is gold ; tell us the good, the bad, and the buggy! 🙏 https://github.com/picselliahq/atlas
Ask us anything below, the team and I will be hanging out here all day (fueled by coffee ☕️)!
Happy Building!
Thibaut & the rest of picsellia Team (@ Picsellia)
