How to and Not to build an AI startup

“AI” as a buzzword has been around in the startup world for a little over 10 years by now.  It started with the early days of OCR system reading address in postal mail and parcel distribution, Email spam prediction on mass scale, multi-language translation systems to name a few. Besides these, all other efforts were mostly research projects and scientific publications.

Two types of businesses and startups have emerged in the rapid growth of AI in business applications and the number of AI businesses.  One, who are building AI-based solutions from the ground up for real customer problems and understand well on how AI solves their customers’ problem. Two, the companies who slap AI and Data Science on their products as a marketing buzzword with a direct repackaging of off the shelf solutions to provide a savvy outlook to their customers.

In my 5 years of work at 6 different startups as a contributor in the core team or as a consultant getting the solutions built from scratch, one truth that stands out is that there is no 1 size fits all solution. In every scenario, be it mobile health imaging in skin problems, human sports performance analysis, robotics or in media entertainment, every off the shelf solution failed miserably and every custom solution solved the customers’ problem in a robust manner.

There is no one-size-fits-all for AI solutions to hard problems.

As a startup founder, having the word “AI” in your startup marketing is inevitably a requirement to sell well or raise money, though the lack of depth of knowledge of your team and the strength of your product stands out in no-time.  If the customers get a whiff of a poorly implemented product or service, your reputation and loss of trust is a permanent loss for your company future.

The first question to ask is, “ is AI the right solution to my problem at hand ?”.  In many scenarios, AI is not the right solution. It is important to remember that AI solutions rely very heavily on data, the algorithms learn from data and it is an established observation that in a well-engineered system, more data yields better-learned models. In almost all scenarios, I have resorted to classical computer vision algorithms and simple machine learning models such as SVMs, Random forest and logistic regression to build the first solutions with small datasets.  It is almost certain to yield a great starter solution with classical algorithms and subsequently train complex models with access to bigger data pools.

AI is data hungry, start small, grow incremental complexity

As a AI researcher and practitioner myself, I share the key factors in the use of AI solutions or building your own AI solutions in your products that I regularly see getting overlooked.

Generalizability: Generalizability implies the ability of the AI system to perform in the general population. Often the data utilized for model development and training is a small sample from the large population. In most startups, the small data vs. big data is a major limitation at the beginning. The models developed in academic research or developed with public datasets are often susceptible to bias towards a subset sample of the population. Incrementally increasing model complexity as the data pool and model learning capabilities improve over time has worked very well for me in all scenarios. 

Robustness: AI models developed to perform with high accuracy and generalizability often are limited by large computation and memory requirements. Their performance in deployment on server/cloud API, edge devices, mobile devices can vary and often suffer from lack of memory, speed or compute power. To achieve robust deployment of AI systems, it is vital to consider the computation, power, network, memory at the design stage of development. Modern methods such as model quantization, knowledge distillation, hardware-aware model training address some of these limitations. It’s almost certain that computation will move to the edge a.k.a. smaller devices.

Traceability: Traceability implies the ability to trace back the steps followed in the process. In the context of AI systems, traceability applies to maintain a record of all model data, designs, experiments and deployment. How the AI systems in the company improve over time with more data and improved model development and training is only possible with strong traceability and reproducibility. In case of a business with strong compliance needs and liability such as finance, healthcare, and utilities, such efforts comply with regulatory needs “for free”.

In summary, all I want to bring forward is that AI as an afterthought in your product development lifecycle is a recipe for disaster, it’s a ticking time bomb that will blow up in the form of loss of customer trust, and risk to your business survival in extreme circumstances.  

On the contrary, AI development and solution integration is done with careful design consideration and thorough and robust development process is bound to reap benefits in due time. It is the single biggest moat that can separate you from your competitors.

The question to ask is “Does your product need or benefit from AI solutions ?” , “ How can you select the best method.”, “How best to strategically commit resources to AI development?”. The key pointers above are essential design considerations, nearly guaranteed to bring the benefits of AI to the product and company.

I advise companies and teams to build AI solutions in computer vision and machine learning solutions. If you do not know where to start to build your AI solution or don’t know how to solve your product problem with AI, happy to answer any queries you may have, you can send me a message on LinkedIn and I will try my best to answer.