Hi
I'm Ayush, a hobbyist researcher, and I like exploring ideas I find interesting in AI. Professionally, I love building tools for researchers and engineers, and getting them adopted. This website has some of my research ideas, philosophy, and the products I've built or helped build over the years.
(Also, this site has a cyberpunk theme you can turn on at the bottom of the page. Try it, it's so fckn cool.)
Backstory
I discovered coding by accident around 2014, in grade 8, mostly building websites, and finally got into ML in my junior year of high school. I've been doing this professionally since 2018, and have had the chance to build some of the world's most popular products from the ground up.
There was a brief period where I considered doing research full time, but my love for building products, talking to users and growing adoption won out over my love for pure research. I've been a student of products and startups ever since.
I generally prefer working at very early stage startups because there is usually less process and bullshit. I've never joined a company with more than 5 to 10 people, and I've never stayed past 250, although I'd like to change that if I start my own.
Currently exploring
Work
Here's a summary of my full-time roles in the industry.
- Weights & Biases. I created and led the open source channel, which ended up being the primary driver of growth for multiple years. I published a lot of popular research pieces, and built partnerships with all the major tools and frameworks, including Hugging Face, spaCy, YOLO and Anyscale.
- Ultralytics. I led the ML team, started the Ultralytics framework and co-authored it, and it turned into the most used open-source computer vision framework there is. I started the community on Discord, and worked on the business side that brought in the company's first revenue.
- LanceDB. I joined as the founding ML engineer, working mostly on AI, devrel and growth. I built the retrieval features of the LanceDB SDK, published research reports and demos, and built partnerships and integrations with popular industry tools and frameworks.
Some of that work you may have bumped into. I co-authored the YOLOv8 model family, led the development of YOLOv5-seg, the first instance-segmentation model to run in real time on a CPU, and software I've helped build gets downloaded millions of times a week. The papers and releases have collected 20,000+ citations.
Blogs (technical and non-technical)
Research
The independent research is mostly multimodal representations, world models, and architectures that are not just bigger transformers. Off hours it's philosophy books and more music than is reasonable.
Talks & Videos
I sometimes talk about what I build, and about AI research in general: conference talks on retrieval and reranking, and 30+ videos taking ML systems apart. A course lesson I taught reached about 950 students.
This site is plain HTML, built by custom agents that pull together what's publicly available about me on the web, and it keeps itself up to date. If you're curious about the setup, reach out.