Minimum Viable Product
You’re an entrepreneur and you need to build the first version of your kick ass new AI startup. But you’re having a tough time attracting a technical cofounder to your startup. You believe in your vision, and you have obtained the money to put behind. If you can get it launched and get some traction, you know that the talent will start coming to you.
Electric Brain isn’t going to be your cofounder, but we can help you get to a minimum viable product.
The typical process for building an AI solution goes in 6 stages:
1 - Design
During the initial stage, we design the full solution. This is often done as a separate engagement, so that you don’t commit all your resources upfront. For certain well known use-cases, we are comfortable providing a full quote without a design session. Nevertheless, at the start of the project, we will go through some design to ensure that we have understood everything correctly.
Please see Service: AI Design for more details on design.
2 - Build the Annotator
An Annotator is a specially designed user-interface which can be used for training AI systems. If there are no publicly available datasets which can be used for your product, we are going to have to gather that dataset ourselves, and we use Annotator to do so. An example of an annotator used for data extraction is here:
We may also build your larger product during this phase, or collaborate with your other development team.
3 - Collect the dataset
During this phase, we build up the dataset that is used for your AI system. Sometimes this entails putting your product into production and collecting data from real users. Sometimes this means a lot of manual work to collect data and feed it into the annotator.
Sometimes this is done by your in-house team. Often times this is out-sourced to third-party companies. For a variety of reasons, we recommend to be careful of out-sourcing firms unless there is a compelling reason to use them.
4 - Research the Algorithm
During this phase, Electric Brain’s AI researchers will go to town trying to crack your dataset. Depending on how much data you had at the outset of the project and how quickly your collecting data from Phase 3, we can either start this very shortly after completing Phase 1, or we will have to perform many months of data collection before the research can start.
This part of the project involves using a lot of high performance servers. If the data you are processing is images, video, or very long documents, this can get quite costly.
Additionally, we can trade off between different levels of quality in this part of the project. We can do a “quick and dirty” research cycle just to get something online, or we can do an extended cycle testing hundreds or even thousands of variants.
5 - Integrate the Algorithm
With your algorithm in hand, its now time to integrate it into your larger product. This may involve synchronizing with your existing development team, or it may involve modifying your product ourselves. If this is a new product, we may build substantial portions of the product during this stage.
6 - Launch, Monitor and further Train the Solution
Now finally you can launch the solution. But it doesn’t end there - many AI providers do not emphasize enough the requirement of monitoring and training. Its very important to use real world, production data in your models. AI solutions are not likely traditional software - you can not just “build it and forget it” like you can with regular software. AI based solutions require continuous monitoring and improvement in order to adapt to a fast paced, changing world.
Time & Cost
Minimum viable products can be quite variable in terms of what they cost and how long it will take to build them. Generally, if you are willing to sacrifice on the scope and accuracy in order to hit certain timelines and budgets, then your MVP can be built quickly and easily.
Cost: $7,500 to $15,000
Time: 3-6 weeks
Depends on: Level of detail
See: AI Design Sprint
Build the Annotator
Cost: $30,000 to $45,000
Time: 8-12 weeks
Depends on: Programming language, framework, complexity of problem
Collect the data
Cost: $5,000 to $100,000
Time: 2 weeks to 1 year
Depends on: Amount of data, time for each unit of data
Research the algorithm
Cost: $15,000 to $75,000
Time: 6 weeks to 20 weeks
Depends on: Expected quality / accuracy level, amount of data, difficulty of AI problem
Integrate the algorithm
Cost: $7,500 to $15,000
Time: 2 - 6 weeks
Depends on: Which technologies we are integrating with
Cost: $2,500 to $4,500 per month
Depends on: Tightness of SLA’s, difficulty of AI problem