I tweeted a thread[0] addressing how we do it. The evolution is as follows: I joined the company when it was a tiny consulting team [as employee number 4] building custom machine learning products for very large organizations. i.e: from problem definition, to data acquisition, model building, application writing that allows using these models, and even allows their ___domain experts to train models with new data.
Now, these projects take a toll on the team, especially when you have people working on different projects simultaneously. The consulting mode has worked for seven years in which the company delivered many projects in many sectors using many techniques. [energy, transportation, employment, telcos, banking, retail, advertising, communication, etc.]
After doing that, certain patterns emerged as we hit road bumps and learned lessons the hard way when it comes to machine learning projects in the real world, with actual stakes. This drove us to start building a machine learning platform[1][2] that takes away the overhead and enables a small team to ship product, deliver value, and do it fast.
We build upon the knowledge we acquired these years and build this in the platform. For example, we enable automatic model/params/metrics tracking and one click deployments because the cognitive load on our data scientists to track experiments was huge, and they didn't necessarily remember to do it, or didn't do it in a similar way. They also had to ask someone who could deploy a model to deploy their models, and this person could be working on something else [bottleneck, social relation].
As we are building this, we also interact with our clients and prospects, some of which are at the leading edge of machine learning and have internal teams, but are suffering from these problems.
So we're working on this to:
- Enable our consulting "arm" to deliver these projects fast
- Enable other people to do that as if they had a team, reducing the barrier to entry as these lessons were learned the hard way [time and money].
Any of the projects we already have shipped could be abstracted and offered as a SaaS to other similar companies in a sector. For example, customer churn for a telecom company. Market forecast. Next best offer. We're choosing to focus on the platform for now, but you can easily see how you could do it were you to choose one project your were paid for and abstract it to other clients.
One important thing is: keeping the conversation open with these organizations. Starting small with a contained specific problem just to get in, and then expanding from that small specific business use case either to expand the tool's capabilities, or to offer it as a SaaS to _other_ customers.
Now, these projects take a toll on the team, especially when you have people working on different projects simultaneously. The consulting mode has worked for seven years in which the company delivered many projects in many sectors using many techniques. [energy, transportation, employment, telcos, banking, retail, advertising, communication, etc.]
After doing that, certain patterns emerged as we hit road bumps and learned lessons the hard way when it comes to machine learning projects in the real world, with actual stakes. This drove us to start building a machine learning platform[1][2] that takes away the overhead and enables a small team to ship product, deliver value, and do it fast.
We build upon the knowledge we acquired these years and build this in the platform. For example, we enable automatic model/params/metrics tracking and one click deployments because the cognitive load on our data scientists to track experiments was huge, and they didn't necessarily remember to do it, or didn't do it in a similar way. They also had to ask someone who could deploy a model to deploy their models, and this person could be working on something else [bottleneck, social relation].
As we are building this, we also interact with our clients and prospects, some of which are at the leading edge of machine learning and have internal teams, but are suffering from these problems.
So we're working on this to:
- Enable our consulting "arm" to deliver these projects fast
- Enable other people to do that as if they had a team, reducing the barrier to entry as these lessons were learned the hard way [time and money].
Any of the projects we already have shipped could be abstracted and offered as a SaaS to other similar companies in a sector. For example, customer churn for a telecom company. Market forecast. Next best offer. We're choosing to focus on the platform for now, but you can easily see how you could do it were you to choose one project your were paid for and abstract it to other clients.
One important thing is: keeping the conversation open with these organizations. Starting small with a contained specific problem just to get in, and then expanding from that small specific business use case either to expand the tool's capabilities, or to offer it as a SaaS to _other_ customers.
- [0]: https://twitter.com/jugurthahadjar/status/131066829330549965...
- [1]: https://iko.ai
- [2]: https://www.reddit.com/r/learnmachinelearning/comments/je0pm...