5 Ways To Make Your AI Project A Success

Don't Panic!

Matt Squire
Matt Squire
Co-Founder of Fuzzy Labs
Mar 8, 2021 5 min read

If you are about to start on a new AI project or are thinking about your first one, don’t panic!

It’s not as scary as it sounds, I promise.

For a start, a lot of the challenges that you encounter on AI projects are the same challenges that you will encounter on any modern software project.

Treat it like a software project

Like any software project, you have to make it applicable to large numbers of people. It needs to be scaled, maintained, and you need to make it open to a large team of people.

There can’t just be one person who’s the expert on the AI project, who everyone follows blindly. The project needs to be accessible so that everyone can collaborate on it.

Also, just like a software project, machine learning and artificial intelligence is a constantly moving target. It’s an area where there’s always something brand new, so keeping up to date with every new change is impossible.

The thing is, AI projects and software projects are heavily intertwined, and to run one successfully, you need to have an understanding of both worlds.

For example, a lot of companies have big data science teams, who are great at building something that shows a concept and creating a model, but the model on its own is of no use to anyone. You need to be able to actually integrate the model into a consumer-facing product or business process. The model is just one piece of the puzzle. You still have to build all of this other stuff around it, which is where the software side comes in.

So, don’t think that an AI project is about moving into completely uncharted waters. It’s not. To help you out, here are my top five tips to ensure that your AI project is a success.

1. Start small

In an AI project, it’s not necessary to solve all the problems at once, so don’t try to boil the ocean. (Not that you would want to boil the ocean anyway… I never really understood that expression!)

Yeah, a flame thrower should do the trick.Yeah, a flame thrower should do the trick.

Build a simple proof of concept that solves one small problem, and build on it from there.

The best way to do this is to figure out what you can build that demonstrates how AI can do the following:

  • Something novel.

  • Something interesting.

  • Something of commercial value.

Once you have figured out what that is, build it in a way that requires as little extra development effort as possible using what’s already available (more on this in a minute).

Also, build it cheap. All you’re trying to do here is demonstrate that something could be done and that there could be some commercial value in what you are building.

Then you can focus on the next steps once your proof of concept has demonstrated what’s possible.

And you know what? If it doesn’t work out, that’s fine too. You can try something different without losing a lot of time and money solving every problem.

2. Use the cloud

In the previous section, I mentioned using what’s already available. What I meant by this is using big cloud providers like Amazon, Google, Microsoft, etc. They all provide off-the-shelf artificial intelligence services that you can pick up and run with straight away.

There are a lot of Clouds to choose from.There are a lot of Clouds to choose from.

These days, you don’t need to build that much from scratch at all. You can get a lot out of connecting existing AI capabilities together to form your own solutions — it’s called AI as a service (AIASS) or “pre-trained modules”.

For anyone starting an AI project, I’d highly recommend using as much as you can from these providers.

3. Understand the data

Binary fully comprehended.Binary fully comprehended.

There are many applications where you would look to use somebody else’s model in your product.

For example, if you wanted to understand customer sentiment, you could train your own model, but that would be a waste of time because it’s already been done.

As I mentioned earlier, all major cloud providers offer this kind of thing as a service. You don’t need to get new data and train a new model unless what you’re trying to do is novel, more bespoke.

Our wound classification project is a good example of this, but it could just as well be a service that labels dogs and cats.

If it’s a more bespoke application and you are using your own data to build a model, you need to first understand what’s in that data and if it’s even usable for your purpose.

Alongside all of that, you also need to think about how bias might affect things.

Let me share a short story.

A few years back, Amazon had trained an AI model to decide whether to progress with a CV or not in order to speed up their hiring process.

So, they took their existing database of CVs, looked at whether they were successful candidates or not, and they trained a model on that basis to reject or accept a CV.

The trouble was this model rejected women much more than it rejected men!

It turns out that the data they trained it with was from all of the male applicants to Amazon, because Amazon’s tech team was overwhelmingly male…

To avoid biases in the data, you have to first know how to recognise them, and that’s a problem in itself. If your entire team is composed of the same kind of person, then there might be biases that nobody thinks to check for and that could end up compromising your data, and therefore the whole project!

Make sure you are taking the time to understand your data (if you are using your own).

4. It’s still software

Many people tend to be intimidated by AI projects, but it should come as some comfort to know that all of the normal best practices that the industry has adopted when it comes to producing and scaling software, apply here as well.

In many ways, it’s run in the same way.

We still need to use version control, we still need to use CI/CD, and we still need to represent our infrastructure as code.

Things still need to be repeatable, maintainable, and so forth. Only, instead of DevOps, it’s data DevOps.

Data DevOps means being able to deal with your data and models in a repeatable and standardised way.

Essentially, data DevOps informs you that your model is telling you* X *because this is the data that was used to train it at this specific time — make sense?

As long as you can understand and apply best software practices to your AI project, you shouldn’t have to veer too far off course.

5. Call Fuzzy Labs 😉

This is far from a sales pitch, but if you want to start off an AI project in the right way, with a simple, easy-to-build proof of concept, then give us a call!

Your friendly team of AI specialists, ready to help.Your friendly team of AI specialists, ready to help.

At the very least, we’d be happy to point you in the right direction.

P.S. For more great content around AI and how it’s impacting the world today, please feel free to email any questions to me at matt@fuzzylabs.ai and follow us on LinkedIn for updates on all our latest projects.

Thanks.


Comments