We Ran an Internal Deep Learning Competition And This Is What Happened

Jonathan Anthony
adappt Intelligence
4 min readNov 3, 2016


Like most companies we have big logs of every site ever visited by our staff. Big brother or common sense is a different debate -“guns don’t kill people rappers do”.

The prize was financial and of course a lot of kudos. These were the rules:

  1. Out of work hours only.
  2. Flexible teams
  3. Two week deadline.

Our aim was to train a neural network which can accurately predict the departments / roles of people based on website logs.

And we had some mind blowing results!

  1. Senior management did not realize that they had given the development teams cart-blanche to expose their online Shopping and Social Networking during office hours.
  2. The Neural network turned out to be able to accurately classify the majority of people by department with over 99% accuracy and with a blind test of humans doing the same the accuracy was less than 50%
  3. The anomalies (people who did not fit the expected patterns of behaviour) were more revealing than the matches.

Results from our teams

Team one : Classified the data with high accuracy and represented it using radial histograms.

Vector based website classification
Neural Network based classification of people
Top sites by person

From team 1, only two people were incorrectly classified. The first person, it turned out the neural network was right and she was actually working as an analyst although we thought she was doing QA. The second person was out of tasks and hiding quietly. We had not realized, but the neural network did!

Team two: Created some of the most beautiful representations of the data and their classification network was both highly accurate and easy to read.

Team three: Created the easiest to follow information and had the most accurate Neural Net.

User activity per hour
User groups, eg. Job seekers
Department classification

They also shone the brightest light on behaviour such as start and end of the active day and browsing of none work related sites. Which underlines the point it is not just managers who care who is putting the hours in. Hard working staff are also very conscious of who is and who is not pulling their weight — especially regarding senior management.

Team four: Took a long range view of the data, which was a real eye opener vs a simple 3 day snapshot and created a simple intuitive way to understand the information.

Team five: has never worked on AI before but were mobile experts and decided to turn their hands to this competition.

The results were beyond what any of us could have expected because they moved outside of the box in their representations

We were left us with an impossible choice for judging the winners.

At the end of this, what is fascinating is the use of AI to discover who is not behaving as we expect, which can have many reasons

  1. Over achievers

2. Under achiever

3. Working / acting outside their remit (malicious / unintentional)

The next challenge for the team is scaling this, from being able to track a small company to being able to track and represent tens or even hundreds of thousands of staff, and in a way that is still easy to visualise and drill down.

As companies grow it is not possible to know everyone individually, but using a dedicated neural network for that task can provide a powerful solution to catching problems early, identifying best potential and visualising staff.

Some final notes, the teams will share a bonus from any future sales. Hats off and credits to

TEAM — 1 : Safi , Jon
TEAM — 2 : Manish, JJ
TEAM — 3 : Senthil, Sankar, Naren, Angu, Mathan, Prakash, Lochana
TEAM — 4 : Sridar, Rimz, Abdul, Nandini, Mohan, Anu, Kalai
TEAM — 5 : Arun, DJ, Rafi, Sheela, Jithan, Sreenath, Rajavali

We are adappt ltd, a Software and Mobile development company with teams in England, India and Spain. Hit the green ❤ if this was an interesting read!



Jonathan Anthony
adappt Intelligence

Software Architect, TensorFlow, Behavioural Analytics, iPhone, Android , TV Studios, Broadcast Playout.