Blog, Events & News

BACK TO ALL

How The Machine Created the Netacea Brand

By Jen Talbot & Richard Jones / 08th Jul 2018

How the Machine Created Its own Name

How we used neural nets and Markov probability chains to create our brand.

One of the hardest things you can ever do in product marketing is to come up with a new brand name. One of the main difficulties is just finding a .com address that hasn't been taken. This alone is next to impossible today without ponying up unconscionable amounts of cash for a cyber-squatted domain.

It's no surprise that so many start-ups have gone with new domains such as the trendy .io and .ai domains. Finding a brand name that we all liked, that was also available as a .com, .net, .ai etc. and had no trademark issues or conflicts, felt like searching for a needle in a haystack.

That's when the penny dropped. Machine learning is great at the needle in the haystack thing, isn't it?

We set off determined to drink our champagne, and design a machine learning program that would accomplish the following pretty ambitious set of goals for the brand creation:

  • Create a brand that would resonate with our target audience
  • Embed what we actually do into the brand DNA
  • Starting from the Engineering team and building our core methods and values into the brand
  • Produce an integrated package of excellent engineering supported by great design
  • Tap into the zeitgeist of the AI / ML learning knowledge base
  • Search through whois and a data registrar to determine suitable available candidates for the domain name
  • Check for any ™ conflicts

"How hard can it be?"

It's fair to say that the prospect of our machine learning algorithm creating its own unique brand name by neatly capturing our zeitgeist, satisfying the complex requirements of our marketing, product, and management teams, performing a search to ensure the name wasn't a potential IP conflict, and finding and searching for available .com addresses, was treated with a great deal of healthy skepticism at best, settling down for a healthy broad-spectrum burst of outright ridicule.

Undeterred our plucky data science team got their heads down and started to look at the best data models, secure in the knowledge that only science could save us at this point.

The first stage of the process was relatively straightforward. We had a team meeting to tease out our aims and aspirations for the new brand via a series of classical brand exercises. We use the five why's if you're brand was a person, and other techniques to build up our core set of brand attributes and brand values. From there we developed a simple spreadsheet of the core values prioritised by the team.

At this stage we would then typically combine the unique values into sets of prefixes and suffixes, and start to combine them manually in unique ways to create unique words, that nevertheless carried through the initial meanings of the brand values we wanted to convey in the root of the word. These combinations are often made easier by including Latin roots, or other languages to create unique words from the combinations, that are nevertheless intelligible and not completely 'foreign'. Typically this is a painful process to do manually and only works with a limited set of values.

We also wanted to tap into our engineering zeitgeist, and expand the base of the possible word combinations.

What better way than just to take the Slack channel conversations and throw them into the data set? So we added our own Slack Channel conversations, our aspirational ambitions for the brand, along with some AI and ML white papers which framed our epistemic knowledge base, into one neat data set.

We quickly identified two data models, the Markov probability chain and the recursive neural net (RNN) which could help us create the prefix and suffix combinations from the knowledge base that we wanted.

Once we had the unique names, we wrote a quick script to check whois for available domains, and a quick search look up to check for potential ™ conflicts.

Initially, we anticipated that the RNN approach would be more useful. And it was, sort of.

The RNN model quickly got to grips with the data set and even started to find us available .com domains. Perhaps over inspired by our Slack channel, the model turned out to be more than a little sweary. We still think Diküm.com is a great choice for someone out there, but on balance we decided to pass on the recursive neural net model and sadly said goodbye to Diküm.com, and switched to the Markov probability chain.

If you want to look at building your own RNN, there is a great tutorial here, it may work for you!

https://pytorch.org/tutorials/...

For the Markov probability matrix, we took the same data set - i.e. all the records from our Slack chat logs, and combined them with our aspirational keywords and used this data set to find our most probable combination of letters. The Markov chain is most widely known as the predictive text used on mobile apps. So for example, if the initial letter was A, the probability of the next letter being B was X. We then generated our probability matrix. We wanted to retain as much of the original brand value and meaning in the original word, and so looked at chunks of 3 letters 4 letters to effectively combine a prefix and suffix with the underlying word.

"No Really. How hard can it be?"

The very first time we ran the new Markov model, one of the very first words, with an associated available domain was Netacea.com. Combining network with panacea, the machine had truly created a unique word. We also quickly found out that the unique word was available for all the major domains and had no IP conflicts.

In order to get to this name, the machine had worked through over 400,000 different combinations.

We had found the needle in the haystack.

As a brand, we wanted Netacea.com to set itself apart and differentiate itself - to have a different outlook, a different approach, that was better than the old generation of rules-based tools, confident and bold - always agile, intelligent, forward-thinking, brave and constantly re-learning.

We do believe that someday, hopefully in the not so distant future, all major websites will use machine learning and AI to manage their visitor flow, and that truly our brand will become a panacea for networks everywhere. Learn more about our Machine Learning approach by downloading the Machine Learning for Account Takeover datasheet.