Everyone is excited about building an AI for everything. Truth is - everyone is far from reality. The best way to build good and reliable products is in fact avoiding the use of machine learning if you can. The best products are simple and just get the job done. Now, if you need ML to do the task at hand, you are up for a challenge and this article is for you.
When does it make sense to use AI or machine learning?
- If a rule based system cannot solve the problem at hand.
- If the problem at hand is worth solving grosso modo (time and budget wise).
- If for <= development cost, you can hire one person to solve it.
- If the best ML engineer you know is 100% confident that the problem at hand is solvable, and Soon.
Else: You will be embarking on a research project that might never see the light of day. Sounds exciting? Well, not for your boss or your CFO.
How do you run ML projects in a real world of budgets and deadlines?
- The product / project manager should be technical. If you are not technical, it will be hard to understand where things are at. Not necessarily but way easier if they are.
- The regular sprint / agile technique won’t work too well. You are not sure what to expect when it comes to adopting new approaches. You have assumptions that some things might work, but they don’t always go according to plan. There is a lot of trial and error involved - making budgeting and a clear timeline challenging.
- Less people, More experience. Instead of hiring 2 senior ML engineers, get one who built a system as similar as possible to yours and pay them twice as much so they stick around.
- Set some budget aside for cloud instances and training data. At scale these make Tiffany's look like Walmart.
How do you hire for ML projects?
- Don’t count grad school projects as experience.
- Look for experience implementing a very close approach in a different field or ideally in yours.
- Don’t hire a data scientist, hire a software engineer. In a high paced environment there is rarely the time for someone new to coding to learn how to implement new libraries, you need people who know how to build things in their sleep.
- Everyone should be patient. The new kid on the block that wants to do 100 things at the same time, will probably ruin your project in 100 different ways. You need someone who patiently tries one hard thing at a time. If their patience drives you nuts, they are probably the one.
Meeting time?
- No more than 2 meetings per week for senior developers.
- Make sure everyone has enough time to come back with updates and results to study.
- Make sure everyone doesn’t have enough time to slack off for more than ½ a day.
This is pretty much how I run ML projects at auxi.ai. These came from a lot of trial and error across the last 2 years, working on a very complex technical challenge while catering for the world’s top consulting firms as clients.
So what did we end up building and where is the ML in this?
auxi is an add-in to PowerPoint that helps you build slides twice as fast. We try to make all features simple to use, but some have loads of technical complexity behind them. Let’s talk about the ML heavy ones.
The smart bar
The smart bar is a command bar that contains hundreds of formatting functions to choose from. This makes the UX of PPT a lot better and more enjoyable. It feels like playing a musical instrument. Now where is the AI / ML in this?
- If you select any items on your slide, the smart bar recommends editing functions to use based on what you have highlighted.
- After you execute some editing function, the smart bar suggests a follow up function to execute and make your overall slide look nicer.
AI editing
Let’s say you are using an old template that has 4 major points. Each of these points has icons, images, charts and many other items that are not necessarily linked to each other. If you would like to add a 5th element to a slide already populated you will have to find or create the space to do it and then make sure that all the elements including the new ones fit well. AI editing lets you just select the items you are interested in, and edit them in bulk:
- Add columns - adds a column to your elements based on your selection.
- Add rows - adds a row to your elements based on your selection.
- Add key takeaways, finds the best place and executes the spacing edits to fit a takeaway message covering your selected items.
AI recommendations
Ever opened Netflix and got the perfect show recommended? This is close but for slides. If you are working on a skeleton slide, where the content is there but the design and branding still need a lot of work you can use recommendations to get design suggestions based on your and your previous team’s activity and templates. It will look at the content you have on your slide, and use ML to recommend the best 10 designs that would work for you.
Then once you select one of them, it will automatically map the current content you have to the chosen design by looking at previous user activity and context of the slide you are going from.
These are some features that use Machine Learning in the auxi add-in, we are constantly editing and working on new products and features.
All in all, it is pretty exciting to be a software engineer in such times. Every week new tech comes out and new possibilities unfold. However, it is very important to keep your feet on the ground, and focus on projects that are not only feasible but also worth working on long term.
Feel free to reach out to me on LinkedIn if you have any questions or comments about this article. https://www.linkedin.com/in/rami-khoury-394820140/
