Peter Yang | January 6, 2020
Artificial Intelligence technology will continue to boom in the coming decade and nothing will stop AI start-ups from improving productivity in every industry. AI products and service will (probably already have) become a commodity that companies HAVE to adapt to stay competitive in their business. This presents huge opportunities for AI start-ups that offer innovative solutions for industries.
Since I founded Awakening Vector, I have witnessed the success of dozens of AI start-ups in North America and European countries, building innovative AI solutions from Zero to One. In this article, I want to share some of my observations, with you – the entrepreneur, in the format of a step-by-step guide on building a scalable AI product.
Step1: Define Your Product
As with all innovative products, the first step in AI product design is to set a clear business goal for the product.
Try to answer these questions to your product team:
- Who is this product designed for?
- What pain points does my product solve for the target users?
- Why are Artificial Intelligence technologies important in this solution?
- What other technologies are needed to complete my product?
- What features make my product stands out?
- What results will my product bring for the target users?
In answering these key questions, two things must be done here:
- Create a case study to visualize the use case and value of your product. You can provide analysis and evidence on increased revenue, reduced cost, improved productivity, and improved performance for your target customer.
- Demonstrate why AI technology is needed in this particular use case by identifying specific functions and tasks in the process that need to be managed using AI technology.
Step 2: Develop Success Metrics for Your Product
With a clear definition of your product in terms of value and function, the next step is to develop an evaluation index for the product to define the success of your product. This step is critical and needs to be completed at an early stage of product development.
A meaningful set of metrics evaluates the business value of the product, rather than the technical performance of the product or AI algorithm. For example, in the case of AI-powered customer service chatbot, meaningful success metrics could be increased customer satisfaction score, increased re-purchase, savings on the cost of labor, etc., rather than solely the accuracy the Q&A threads.
Besides, the success metrics should be easy to measure in dollar value for the target users of your AI product.
Step 3: Develop Your Data Strategy
Data is at the core of AI products. If creating products is like building a skyscraper, data is like the bricks used to build it. You might think there are more data in the world than you will possibly need. It is true. However, do you have access to the data your model needs? Can you afford the data?
Once you decided what type of data you need and figured out where to get them, your next task is to acquire training data for your model. This might be the most critical step in developing and improving your AI model, as the quality of your training data will determine the accuracy of your AI model, hence the performance of your product.
While some data are public and free, they are unlikely to fit your model exactly. You can get these data from a public database or with a crawler. Be very careful about the potential, and frequently hidden, the bias in the free data set. If your training data is biased, your model will be as well.
Another potential risk of free data is legal risk. For example, some data set may contain a large amount of personal information, which, if misused or leaked, could lead to serious legal issues for product developers. In recent years, there are many start-ups failed due to improper handling of sensitive data.
The best way to get quality training data is by using a customized data annotation service. The industry has involved drastically in the past few years and the cost has gone down significantly. Awakening Vector provides a fast and low-cost manual annotation service with over 99% accuracy.
Before you hire a data annotation service, think through the following questions:
- How much data is needed to train your model?
- How to classify and organize the data?
- What’s the proportion of the test data to the training data?
- Do I have a detailed description of the labeling tasks or problems?
Step 4: Design Your Product Model (ML)
AI algorithm and model design is the technical core components of your product, which are usually the responsibility of developers. Therefore, product developers are required to have an in-depth understanding of different models and AI algorithms to select a useful model for the product based on the market data. You can build a new model by yourself, or use off-the-shelf resources like AutoML.
Below is a quick comparison between build-your-own vs use existing MLs:
|Build-Your-Own ML||· No restriction on usage scenarios
· Fully controlled parameters
· Full customization
|· High cost
· Rely on model algorithm experts
· Limited external support
|Use Existing ML||· Stable and robust
· Easy and fast to operate
· Less costly to develop
|· Limited use scenarios
· Difficult to scale
· Data cannot be privatized for deployment
Pay attention to these three steps when you are designing a model to build one that fits specifically to your software solution:
1) The selection of activation function
2) The setting of training weights
3) The use of node types and structures
It is important to note that the test data and the training data should be separated. The test data can be put in use only after the training is completed on a separate set of training data. Common test indicators include accuracy, recall rate, F1 score, and obfuscation matrix.
Step 5: Build Your MVP
The leaner, the better. Building a Minimal Viable Product (MVP) is at the essence of building a lean product. If you are not familiar with the lean concept, I recommend you start with The Lean Startup. This is an excellent book for tech start-up founders.
Prototyping is the core process of building your MVP, and you will need many iterations of until you get the MVP that your potential customer likes. For software products, you can test the interaction with UI mockups; for hardware products, you can make a concept video through 3D modeling software without having to build it in a factory.
At this stage, it is also important to start building your brand, awareness in the community and start to establish sales channels.
Step 6: Growth Hack
Your final step in product development is to build a growth loop that enables long-term growth. Use proven growth hack techniques such as user behavior tracking, A/B testing and in-app incentives to do growth-hack. Hacking Growth by Sean Ellis is an excellent book to learn real-world examples of successful growth hacking strategies.
As your customer base grow, you will have more data on how your customers use your product and insights on where to improve provided by your customers (not assumed by your product lead). Please continue to monitor the performance of your product and continue to improve user experience.
The Last Word
Now you have walked through the six essential steps to build your AI product, does it feel difficult? Not really! Building an AI product is not that different from building any other technology solutions. The key is simple: solve a problem.
In my opinion, the design of AI products is a process of turning imagination into reality, imagining an ideal solution to a problem, splitting the solution imagined in mind into small tasks, and transforming these small abstract tasks into data annotation tasks. After that, it’s about choosing the right data set, choosing the right model algorithm, and training your magic solution step by step through trial and error.
About the Author
Peter Yang, Founder & CEO
Peter Yang, the man behind Awakening Vector. After witnessing experiencing the headache of labor-intensive and costly manual annotation work, in 2018, he gave up a career in VC to start his own company, a data service provider trying to build a one-stop data annotation for AI companies worldwide.