What is Iskra?
Iskra is a zero configuration platform for machine learning. We provide storage and management for your datasets, infrastructure for training, and can deploy your models to web (well, that's coming soon!), all with just one command typed in the terminal.
In practice, it means that with Iskra, you and your team will never have to worry about: configuring cloud providers, setting up Kubernetes, installing dependencies for training, choosing optimal hardware, configuring network for deployment of your models, setting up certificates, setting up CI/CD... the list will grow with every feature we add.
You can see our development roadmap down below, and are more than welcome to contribute to it - feel free to drop us an e-mail or open a chat in the right bottom corner of this page. We love feature requests.
What is the motivation?
We believe that it takes simply too much time to develop any machine learning, and especially deep learning based projects. According to Algorithmia, it takes 3 months on average to deploy a single model.
We see that there are plenty of tools that accelerate this process, often to just a few minutes, but they all are either incomplete, or just not enough to use on scale. Oftentimes you have to adapt your code, give up existing tools, or, especially, deal with problems that generally don't belong to your area of expertise just go get basic things running. Our philosophy is to make Iskra a platform that:
- Doesn't require you to change your code
- Requires no prior learning to be used
- Requires no maintenance
- Doesn't increase costs
- Covers the whole machine learning workflow
We are an early stage, VC-backed startup and more than anything, we're looking for as much feedback from you, ML experts, developers, redditors and humans of all kind - let us know what you think!
For now, we run fast (and hopefully don't break too many things), and that means that our planning is mostly short-term. Having said this, we know about a few key features we will release for sure in the forseeable future:
- Deployments to web within one command
- On premise usage - use our platform with your cloud provider
- Realtime exploration of training artifacts
- Experiment comparison (potentially, integration with one of existing tools)
Of course, there will be a lot of smaller, platform related improvements coming soon, too:
- E-mail notifications for trainings
- Web interface for dataset management
- Webhooks of all sort
If you feel like there is something obvious and crucial missing on this list, let us now. The tiny chat bubble on the right bottom corner is waiting to be clicked!