Dr Sundeep Teki is an AI Leader and Consultant. Previously, he conducted cutting-edge applied AI research at Amazon Alexa AI in Seattle and Swiggy in India. 

Sundeep obtained his PhD in Neuroscience from University College London and a Master's in Neuroscience from Oxford University. He also organizes a course on accelerating data science careers for working professionals.

INDIAai interviewed Sundeep to get his perspective on AI.

How crucial is starting early for those interested in data science?

It is essential to build a solid knowledge base which can take time. Therefore, make this foundation earlier than others. Then, you can advance on the journey faster and develop better first-principles thinking and intuition for various machine learning problems. 

However, regardless of when you start your career in data science, the key is to keep practising and honing your skills, given that the field of data science will continue evolving rapidly. I have worked with bachelor's students and senior IT professionals in their 30s and 40s who are equally motivated to kickstart their careers in data science. Any motivated person can become a good data scientist with many open-source resources, courses, and datasets available online. 

What advice do you have for people who want to work in the field of research in artificial intelligence? What should they focus on to move forward?

Innovative AI research is not straightforward to conduct. Researchers should focus on a problem area in which they are interested. If the research problem area is grounded in the real world, practitioners can test their algorithms on real-world data and improve from the feedback.

The most important skill for doing novel research is to think deeply about a particular problem and apply the scientific method. This skill involves coming up with relevant hypotheses and conducting several experiments using suitable datasets, algorithms etc., to test the validity of the hypotheses.

To come up with innovative ideas, you need to know the existing literature and what ideas have previously worked or not worked for a particular problem. Unfortunately, knowing what ideas can generalize and are practically feasible to solve a business problem is a rare skill that distinguishes the top applied researchers from the rest.

Should startups build their own AI solutions for their software or use AI software made by someone else? Does the price difference come into play? 

The challenge of building vs buying is a very contextual decision. For startups, it depends on several factors: 

• including what stage they are at (early-stage vs growth-stage); 
• whether their core product or service is AI-first or business-driven; 
• whether they have an in-house data science team or not; and 
• whether they have the requisite funding to invest in AI solutions from third-party vendors.

Pricing is an essential factor underlying this decision, as it is usually faster to have trusted vendors build the product than to hire an in-house team and wait for them to ramp up and deliver on the product. Unfortunately, I've seen startups with inexperienced leaders too often take a short-term view based on pricing and sacrifice the quality of the solution. This approach inevitably results in multiple problems, including delays, lack of accountability, and poor or unmet business outcomes.

What inspired you to become an AI Consultant after spending time at big tech in the USA and startups in India?

I've been fortunate to work on several aspects of AI - from working on real neural networks in the brain to market-leading consumer products like Amazon Alexa based on cutting-edge AI. With hands-on experience developing brain-inspired algorithms and translating them into real-world AI applications, I have developed unique approaches to solving various AI problems.

As an AI Consultant, I can leverage my deep domain expertise coupled with a practical understanding of how startups and big tech companies take AI products to market. No two clients or business problems are alike, and collaborating with them to solve their unique technical and organizational challenges is a fulfilling endeavour.

I particularly like working with early-stage startups and helping founders develop an appropriate Data and AI strategy given their particular line of business, scale, and use cases. 

People usually think that AI infrastructure costs a lot of money. Is this statement true?

Yes and no. It depends on who you are.

Building the infrastructure to power the entire machine learning lifecycle from raw data to AI predictions is an expensive yet necessary requirement for an established startup or company that relies heavily on AI for its core business. The cost of the AI infrastructure can be exorbitant, from data warehouses to feature stores to compute and deployment resources. However, if companies can execute their AI strategy and vision properly, such investments are usually profitable in the long term.

If you are an entry-level or experienced AI practitioner, the cost of training AI models and deploying them via apps has significantly come down in the last few years. You can build many AI products using free credits from cloud providers, accessible GPUs from tools like Kaggle or Google Colab, and leverage open-source datasets, algorithms, libraries and pre-trained models.

How hard is it to determine all the costs of AI? What was your experience with this?

Most companies' tech and AI infrastructure are by cloud providers like AWS, Azure or Google Cloud. These cloud providers provide tools that help you estimate and review the costs of various AI tools and services, making it easier to have a transparent view of AI costs. 

On the other hand, if you are a bootstrapped company using various solutions from different sources, keeping tabs on the costs can be a bit more cumbersome.

What are some of the most important things you've learned from this?

Developing and executing an AI project to fruition is a very challenging and complex task. However, you can increase your odds of success by having a systematic process of identifying, prioritizing, and developing AI use cases that can yield a more significant business impact. Unfortunately, most startups do not have a structured approach to building AI projects and blindly apply the same process of building software products to AI. Establishing a clear roadmap and processes, identifying and aligning cross-functional stakeholders, hiring the right team, and having patient and supportive leadership in building a productive AI function are paramount to achieving the business goals.

I have written about this previously in my blog.

Could you recommend some books and articles about artificial intelligence that you know are good?

Here are some recommendations:

• Pattern Recognition and Machine Learning
• Deep Learning 
• How to Create a Mind 
• Vision 
• Books with practical, hands-on components published by O'Reilly