<aside> 🌐 Sieve is the platform to discover, design, and run AI features at scale. AI development is moving away from being predominantly about training & finetuning, and more towards how people combine models together to make specific use cases work. Our platform is the best development experience built around this growing trend.

</aside>

Problems

Concretely, it’s still really difficult to find the right models for a specific use case, combine them with just the right logic & other models, and then deploy that pipeline at scale — especially for team just getting started with AI.

Discover

HuggingFace is amazing, but it’s pretty overwhelming for most people with little AI knowledge. Weird model names, many models that don’t work, unintuitive demos, and maybe too many options to choose from. Also, not every model is on there and some of the most useful models for niche use cases are stored as messy GitHub repositories with weights on Google Drive.

Design

Once you’ve found what’s out there, you’ll design a pipeline and see how it performs. Model code tends to be messy so you first need to “make the code work” with all the weird CUDA & package dependencies. And once you have managed to get things running, you spend a lot of time commenting / uncommenting parts of your code, running steps in isolation, storing intermediate outputs to debug a specific portion of your pipeline, swapping out specific models (like a face detector), writing custom visualization code, and then fixing your pipeline logic. This ends up being a tedious and cumbersome to go through.

Run

After you’re satisfied with your pipeline, you may choose to have others on your team try it. This may mean spinning up a quick UI for your team hosted on a GPU somewhere, which takes some unwanted effort. Once you’ve got the internal pass, you’ll then look to move this to a setup that scales in production — which means worrying about Kubernetes, CUDA drivers, autoscaling, observability, GPU utilization, large batch requests, retries, error handling, and more.

Product

We think there is a better experience to be built around the three steps:

• Discover: A library of relevant solutions, models, and utilities with nice demo interfaces
• Design: Code-first and visual experiences around function outputs and pipelines.
• Run: Scalable GPU infrastructure with all the knobs a developer may ever want

Product Strategy

While we believe the core platform experience is the most valuable part of Sieve, it doesn’t mean much unless we prove to people that the platform can easily solve use cases they care about.