Economics. MBA. Investment Banking. Social Development. Travel Entrepreneur. Product Management. This has been my career so far. The first time I heard of deep learning was just over two years ago. Even when I did, I had no clue what it was capable of and how it did what it did.

Today, I am the founder of Agara Labs, a startup developing cutting-edge solutions in customer support using deep learning. We sell to some of the largest enterprises in the world and are pushing the boundaries of applied AI (sometimes behind closed doors).

Not only am I doing well, but I am also actually enjoying the journey more than my previous gigs. These are my learnings.


Let’s start with some facts.

#1 Deep learning is a tool, not a product. You still need a product

#2 Good products ≠ More deep learning. Choose wisely

#3 Clients know little about deep learning. Also, they don’t care

#4 Few investors get deep learning. All investors get good products


The above facts set the stage for what I do every day.

This is not a how-to guide. These are thumb rules I have learned in the last 18 months running Agara Labs. These thumb rules have helped me make sense of things as we have grown and, I believe, will continue to hold true as we progress. This is also not a getting-started guide. These thumb rules assume that you already have a co-founder or employee well versed in deep learning. If you do not, go find your deep learning partner and then read this again.

Read like crazy

Deep learning as a field is so new that even the practitioners of the trade constantly feel being left behind. Thousands of researchers across the globe are working on problems and publishing papers with alarming regularity. The cutting-edge is changing quickly. The only way to stay ahead of the race is to read like crazy.

Read industry publications and semi-research publications around deep learning and how organizations are adopting it in their work processes. Read about experiments being conducted by universities, companies, and governments. Read the Twitter feeds of academics and founders of deep learning companies.

This diagram while initially incomprehensible becomes easily understandable within weeks of getting started. Reading articles like this from MIT Technology Review gives a great theoretical view + practical application of existing as well as emerging tech.

Most technical problems and questions will become obvious when you have read the first 10 articles or books. Some others you will be able to decipher through simple Google searches. When you hit the boundary of your understanding, move to the next step.

Be the stupid guy

When you enter a room full of deep learning scientists, accept that you are the stupidest person there. All of the deep learning guys in my team are younger than me. Most of them are younger by well over a decade. Given the age of deep learning itself, you are going to be faced with a similar setup.

Let go of any inhibitions and ask questions.

Ask how something works. Most likely you will not get an answer that goes in a straight line. Let that not faze you. That is how deep learning operates — its fuzzy. Keep asking.

Ask what you can change in their models. Ask what happens when you change the inputs. Ask what happens when you change parameters. Ask how you determine the parameters. Ask how easy or difficult it is to change any of these and how much time these changes may take. Some of the answers will surprise you.

Ask what you can do to help. Ask what data they need and how much. Ask why. Drill down into exactly what they need and you might be able to offer entirely different suggestions. Sometimes you will surprise them.

Most importantly, though, as the product owner, you need to ask the most important question for which we move to the next step.

Find the break-points

You may never fully understand how the internals of a deep learning model work. You may never be able to convincingly answer questions like ‘Does the model understand this sentence or is it looking for keywords?’. The good thing is you don’t need to.

The precision-recall curve. As you increase recall (try to answer more questions), your precision falls (how correct you are answering those questions). This is similar to human comprehension. You can be great in a limited field of study. As you expand the field of study, your knowledge levels drop overall. Photo courtesy: www.andybeger.com

You must, however, understand the precision-recall curve and know that your product has limitations. You must understand the conditions in which performance starts to deteriorate. You must understand when your product’s failure will start to become obvious. You must know your product’s break-points.

Finding the break-points will never be as easy as asking someone. You will need to test the hell out of your product. Even small changes will require you to re-test everything afresh even if you think they would not have any impact. If you go into productions, be prepared for a flurry of false-positive instances being reported by the users despite your best efforts.

Now that you understand limitations, your natural reaction as a founder (as it should be) would be to break those limitations. Which brings us to our next step.

Double check your expectations

All founders, especially restless founders, constantly try to push their teams to come to their lofty dreams. This is a good thing. However, when it comes to deciding the product you intend to build, it is important that your dreams are matched to what the science can achieve.

One type of risk comes up when you think of deep learning as capable of doing everything — perhaps from having your entire knowledge derived from sensationalist publications. These often come crashing down when the results come nowhere close to expectation and you take out your frustration on your team.

We just hired a smart young deep learning engineer whose previous stint was to develop a generalized summary generation engine to capture small nuances using multiple knowledge sources. This is something that is doable — but will only do well for some domains. The founder of the company, however, insisted on this happening for all domains quickly. He held daily sprints to ensure there was rapid progress being made. All good on paper but in reality, this needs time and focused work to achieve, even in a limited scope. It needs a bunch of experimentation, a lot of which may not yield immediate results or any results at all. It was no wonder that within weeks both the deep learning engineers decided to move on.

An equally big risk is selling yourself short. As a founder, your greatest strength is not assuming anything to be axiomatic — fixed and not available for modification. You must apply rational thought to all challenges and limitations before accepting them as final. Once again, in deep learning, this is a harder problem for a non-technical founder to solve. This is why this step comes after asking you to read extensively, question comprehensively, and make an independent judgment.

The story of Elon Musk and SpaceX always fascinates me in this scenario. Faced with a minimum $20m price tag on cheapest rocket he could find, he delved headlong into the mechanics of building a rocket. He came up with a full bill of materials and established that there is a basis to build a rocket for far cheaper. That was the starting point of the revolution in breaking the monopoly of half a dozen companies in the global space-faring industry.

From the book Elon Musk by Ashlee Vance

Admittedly, the above two risks counter to each other and, ever so often, appear together. The ability to distinguish them and solve for them appropriately is what will decide whether you build a me-too or something truly pathbreaking.

Now that you have a (hopefully!) clearer picture of what you want to build, it’s time to get to the day to day of product building.

Personally run your team standups

Team standups rarely end with major decisions being made on the spot. However, as we saw in ‘Find your break-points’ section, small changes in approach to the models can have major impacts on the precision-recall curve. You need to be in the know.

A lot of times you might find yourself lost in the technicalities. Let them take you over. Push decisions for a day. Spend time after standups to read, ask questions and make up your mind. Then make the decision and monitor what happens. After the first few times, you will find yourself intuitively recognizing the preferred path. Once again, all the reading, questioning and judgment making will come in handy.

Allow for experiments

The rush to production and the pressure to show constant progress will cause you to expect rapid movement all the time. Resist the urge.

Classify what is really important to your product in the next few weeks, what will be good to have in the next 2-4 months and what might represent a great next step 6 months out. Use this framework to prioritize tasks at hand. Move the immediate deliverables in standups. Ensure that there is an ever-growing understanding of mid-term objectives and resources are allocated accordingly. For the long-term objectives, check whether you have the right people, tools, and data. Start gathering and allocating resources but don’t get hassled if you don’t see results week after week.

For Agara Labs, the current set of objectives is clear. This ensures the deep learning team knows it has sufficient time and resources to dedicate to thorny, peculiar problems.

Immediate objectives: Voice Assist v1

Mid-term objectives: Voice Auto v1, Text Auto v1, Text Assist v2 (multiple language support)

Long-term objectives: Voice Assist v2, Voice Auto v2

Double up on engineering

You will quickly realize that deep learning is a small part of whatever you build. Without a robust engineering setup, your innovative solutions will have nowhere to go. The engineering around the core deep learning bits is also easier to understand and optimize on. Go at it with all you got.

You will also realize that as the complexity of your product grows, you will need smarter engineers to simply keep up. Invest in them. Not just in hiring them but also to keep them up to speed with what your deep learning folks are involved in.

Be the customer

Remember fact #3? Your customers have likely heard about deep learning and maybe read an article or two. They may have had a bunch of other founders like you pitching to them about deep learning. They absolutely have been bombarded with context-less articles about how deep learning is either the savior or the devil, depending on their reading habits. However, they do not understand what it is, how it does things and what it can potentially become. You, having read extensively, questioned the good folks and made multiple judgment calls, now have a fairly strong understanding of all of those things. Use it.

Translate the esoteric jargon of academic publications and your lofty technological dreams into simple how-does-this-benefit-you points for the customers. This, more than anything else, will ensure that your deep learning product sees the light of day.

This is also the reason why you — the technically illiterate founder of a deep learning company — are better off than some of the smartest scientists you will meet. You must have the ability to convert technology into benefits. Scientists tend to be too deeply engrossed into what they can do. You, on the other hand, can be focused entirely on what they can do for the client. That is where the battles are won.

Enjoy yourself!

Any startup journey is hard. One where you do not fully understand what you are building is significantly harder. So it is important that you try and find the fun in the middle. You will go wrong. You will make the wrong calls. Your models will fail. Your demos will refuse to work during critical meetings. Laugh and get prepared for the next challenge.