Based on my experience, building a startup is not an easy task. It does require perseverance for a long time. Many of us, especially millennials who are likely to take a leap of faith and jump over onto the battleship without proper planning because thinking that they could make fast money and freedom in terms of time. But the fact is that the reality is totally opposite. 90% of the startups will fail in the first year of the operation and they ended up became a part of this statistic.
Recently, there’s a lot of perceptions from some entrepreneurs towards AI that AI are AWESOME, WORKABLE and TRENDY. According to Marvin Minsky as interviewed in Hal’s Legacy,
Only a small community has concentrated on general intelligence. No one has tried to make a thinking machine. The bottom line is that we really haven’t progressed too far toward a truly intelligent machine. We have collections of dumb specialists in small domains; the true majesty of general intelligence still awaits our attack. We have got to get back to the deepest questions of AI and general intelligence. “General intelligence” does not mean exactly the same thing to all researchers. In fact, it is not a fully well-defined term, and one of the issues raised in the papers contained here is how to define general intelligence in a way that provides maximally useful guidance to practical AI work.
Back to our topic, how do you actually build an AI startup? There are 5 elements you need to consider before building an AI startup:
- Blue ocean strategy
Blue Ocean Strategy
What is Blue Ocean Strategy? According to Wikipedia, Blue Ocean Strategy is a marketing theory from a book published in 2005 which was written by W. Chan Kim and Renée Mauborgne, professors at INSEAD and co-directors of the INSEAD Blue Ocean Strategy Institute.
From my understanding and perspective view, it’s a kind of strategy whereby the capability of the company that systematically created an uncontested market space that automatically makes the competition irrelevant. This type of company has higher product differentiation and lower cost to operate on it.
In order to achieve this, the startup needs to know the most important part; what is their USP (Unique Selling Proposition) compare to other competitors, target market (local, South East Asia, Asia, global?) and gain some tractions to validate market needs (user growth or revenue).
Is the startup focusing on horizontal or vertical problem? For example, IBM provides a solution that can solve general problems such as data analytics, cloud computing, etc. That is an example of a company that tackles for horizontal problem including Deepmind, Amazon, Facebook, Microsoft, and Baidu. Compared to vertical that are more focusing on a specific problem such as Tesla or Uber, they are focusing on the transportation problem.
Like us (a Malaysian startup, Soding), we are running an AI-powered recruitment platform for employers to hire great software developers. What makes us unique compared to other competitors (Codility & Hackerrank) that we have our AI technology to analyze technical skills and personality. But to be highlighted that our target market is SME and tech-based companies which only covered in the South East Asia market with potential revenue RM80 mil. Currently, our traction is the several companies and tech candidates that have already registered with us. Our clients are mainly from local, talents are global but we are still doing product-market fit until now.
To develop an AI product is not an easy thing. Nowadays, there’s a lot of AI startups actually still weak in terms of AI characteristic inside their product. This disadvantage can lead other competitors and ease them to copy out and build for their own. Try to build a minimal strong AI even it’s not perfect as possible in order to achieve uniqueness. You don’t need to build your own AI framework, you can use the existing one such as Scikit-Learn or Tensorflow and your solution must be a focus on the specific customer problem.
Like I mentioned before, most the entrepreneurs with technical background, their perception towards AI must work 100%. If they use human in AI is a failure and it’s a must to be accurate 100% most of the time. All these perceptions are absolutely WRONG.
Let me tell you about our story. Our product does not require to work 100% as long as it can deliver the result and convenience to them. Currently, half of the processes that only can be automated because the AI engine part still requires a lot of training data to feed for predictive modeling to run efficient and accurate thus gain feedforward-feedback between us and clients for product-market fit the purpose. Before this, we joined one of the accelerator programs. One of our mentor curious about why he/she can’t see or use the product yet. It is because the previous reason forces us for not showing the complete product instead of just a green terminal in front of him/her. Our product consists of three parts which are static code, predictive and personality analysis. We used a combination of AI frameworks that will be our main state of the art. Our predictive modeling still needs human intervention in order to make it right, but not only smart. There is a requirement of 100% accurate unless high product risk such as healthcare that needs serious attention on it especially when involving human life. A tiny mistake can cause serious damage or maybe death. One of my audience during an AI event, CEO iflix Malaysia, Azran Osman-Rani asked me, what are the differences between our technology compared to IBM Watson. On that time, we used NLTK, one of tech stack focusing on Natural Language Processing to process social media and candidate data for sentiment analysis (only English language). I replied to him that our technology during those time just below 80% accuracy based on the collected training data but the answer was not really satisfied him. But we do realize during the product-market fit period, there are pros and cons if we are using NLTK and IBM Watson technology but I can’t tell more details about it since we are partnered with IBM. Maybe one day, I can confront him to share our finding.
Moving back to our story, why we still need human even automate? It is because to avoid disaster even full autonomy. User experience for product-market fit to discover a new thing based on the feedback from the market and for sure to train to model in order to be smart and right.
When did we talk about the data, how we can retrieve it to train the AI? It can be in many ways such as existing dataset from database or data crawling using Beautiful Soup or Tweepy. But remember, Malaysia has a privacy law called PDPA (Personal Data Protection Act 2012) to protect your personal data. You need to do some data exploration analysis. For instance, identify some features. After that, clean the data (outlier fixing and impute missing values). It might take 40% of your time. Then use a machine learning approach for modeling. Finally, evaluate your modeling to determine performance such as accuracy. You need to optimize your modeling from time to time in order to make sure that you can get a better result.
When dealing with the clients, DON’T ever requests any data without completing your modeling and you must test it with your OWN dataset in order to show more value-added.
Rule of thumb for your startup is to make sure split into 50/50 for business and product development. It is because you must make sure its balance between these two in order to improve both processes at the same time.
In mid-2016, we were surprised by the local startup scandal whereby the tech guy was ditch over by the group of co-founders. At the end of the story, that startup business process shook up when the tech guy steps out with his tech. Moreover, there is no product whatsoever for their business. Same goes to AI engineers, they are a novelty in each of the startups. So please, do appreciate them.
In business development, the person that handles this process must know some fundamentals of social engineering knowledge. He/she must be able to educate the users (even it’s quite hard and takes a lot of time) in the market, helping clients to solve their problem and also take care of the security and privacy of the stakeholders. For us, we have our own internal QC (quality control) to take any preventive actions instead of corrective, if any future problem occurs.
There’s a lot of misconception about AI term. Many people that self-claim knew about AI always keep abusing AI term as we heard on social media and sharing knowledge event. What they know is AI will be replacing the jobs in the future, destroy humanity and so on. Another thing is that they claim using an AI framework is easy as 1,2,3. Even I’ve graduated my Master Degree in Computer Science and worked in that field since the end of 2015. I never thought that it was simple just like that.
Do you ever hear about ConvNet (Convolutional Neural Networks)? This “thing” specialized in image processing technique. First, you need to prepare a subset from one sample image. Once prepared, set a filter. Convolve the filter with the image like slide over the image spatially, computing dot products. Hence, the results are by taking a dot product between the filter and a small chunk of the image. Prepare a few activation maps and convolve all the slides all over the spatial locations. Below is the code sample for ConvNet using Tensorflow:
What we can conclude from the above explanation, we really need a deeply technical team, especially that have a core skill in machine learning. Most of the time we can get it from Ph.D. students. Also, the top engineers who can produce and deploy AI. Data engineer, scientist and analyst are totally different roles. A data engineer is responsible to make sure clean, prepare and optimize data for consumption for the data scientist to convert the data into storytelling. A data analyst collects the data and helps companies in making better business decisions. But for the startup, you need to find a great talent that can do both in order to save your resources. In most cases, the CEO needs to be good at technical too.
It’s rare to find this kind of talents in the local market because the talents prefer for a stable company that comes with more benefits, likely apply to MNC companies and more secure (that’s why Soding exist!). Tech companies are willing to pay for hiring because it’s hard to attract local talent. My advice when you are seeking for talents, find this kind of traits which are business domain and have technical nor mathematics skills. It’s part of the DNA requirement for team members.
Our team consists of 7 people whereby 2 full time and 5 part-time. All Malaysian except 1 from Portugal.
At the end of this article, I conclude that these 5 elements can help you out in order to build an AI startup. Are you ready for AI technological advancement?