We wanted to have a positive impact on a problem we care about while having a sustainable business. Since we have experience in multiple challenging work environments, we decided to focus on wellbeing at work. Considering burnout is affecting an increasing number of people worldwide, and especially young generations, while being hard to detect early enough personally, we ended up designing an AI-powered tool. Since existing solutions related to wellbeing are mostly dealing with company/team performance, we decided to focus on employees by providing individual and anonymized diagnosis and solutions. Eventually, as employers are customers, we wanted them to be able to track progress and therefore design collective metrics, which gather individual anonymized data.
What it does
We provide a 3-step solution working as a virtuous circle:
- Identify: Detect early signs of burnout on an individual basis, thanks to AI and Data Science, and without requiring any contribution from the employee. This individual Burnout Scoring is the main value of our solution.
- Solve: Provide the employees with customized support once identified as at risk and depending on their Burnout Scoring: recommendation of daily activities for medium score, and connection to well-being professionals for high score, by partnering with already existing solutions.
- Improve: Offer weekly/monthly/annually analysis to companies based on their employee's wellbeing anonymized data, and help them to improve it through a dashboard that provides them with metrics and graphs.
How I built it
We created a prototype of our tool during this hackathon, by focusing on the Burnout Scoring. This first version of the burnout scoring is made of 5 different metrics, based on the scientifically proven behaviors associated with early signs of burnout:
- Demographics: gender and age
- Sentiments: obtained analyzing written conversations with the Google Cloud Natural Language API
- Insomnia: as soon as the person uses his electronic device during sleeping time
- Effectiveness: depending on the number of emails read on a daily basis
- Seasonality: depending on daylight, business cyclicality, main holidays and bonus payments
Challenges I ran into
Define simple metrics to have a first Burnout Scoring easily available without breaching data privacy. Because of lack of open dataset, we created our own using scientific studies to generate a likely-to-be-true dataset. Successfully connecting to the Google Cloud NLP API and preparing the acceptable inputs was the following challenge. From the API we generated the sentiment scores and magnitudes of a series of real tweets simulating employees’ tweets. We then used these results and the other scientific signs of burnout to train our Machine Learning model. We identified Logistic Regression as the best model for our application. With an out-of-sample accuracy of 85% in identifying burnouts computed on the test set, we then confidently created different profiles to be shown in the demo. The integration with the front-end has caused us several issues, but we were able to solve them with the help of GitHub.
Accomplishments that I'm proud of
This was the first hackathon for 2 members of our teams, with no background in computer science nor in data science. Moreover, it was the first time the 4 of us worked together and we barely knew each other. Despite this, we were able to collaborate really well and find a way so that everyone good contribute to the teamwork and learn from each other. Working in parallel allowed us to manage our time effectively and came up with a first prototype of our product.
What I learned
How to use the excellent resources of the Google cloud API and tune them to solve existing problems. How to effectively connect back-end and front-end. How to use GitHub. How to build machine learning models in Python.
What's next for WAW
The next step of our project is to improve the Burnout Scoring by using the most famous prediction Burnout questionnaire: the Maslach Burnout Inventory. It gathers 22 items related to employee wellbeing.
We want to improve:
- Health scoring: For now, we only take into account data about insomnia. The idea is to get the data from Sleeping, Fitness and Health apps in order to establish a scoring, including for example the delta between current and needed number of hours of sleep, or the level of exercise.
- Seasonality scoring: establish a precise scoring of the periods of the year in which people are the more likely to be stressed or depressed depending on bonus time, weather, holidays, etc.
- Professional effectiveness: improve the effectiveness scoring by taking into account the evolution of time that employees are spending to write an email over a few months and analyze the reactivity to messages of their colleagues on Slack.
- External data about the health of the company: Are they on a recovery plan? Have they just signed a meaningful contract?
Then, we will improve the model of prediction we created by training it with real data from an Institute such as the Berkeley-Stanford Computational Culture Lab.
Moreover, we will test our Business Model assumptions starting with our value proposition and customer target by creating an MVP for the users and the customers and asking for feedback about it. We will constantly proceed to discovery interviews in order to adapt the project and understand the users’ and customers’ needs.
Finally, we still have to overcome data privacy issues, both on the employee’s side (personal data) and on employer’s side (customer data).