Greenwood is an NPO that provides career opportunities to Latinx and Black Communities within the financial industry.
It has been 3 years since they implemented their business model of matching students with mentors for guidance in career prep. Having witnessed success with over 300 students in 2019, they are now looking to expand their reach to the national level. With this expansion, they expect their students to increase by an exponential amount.
Currently, instead of working on their normal duties, employees of Greenwood have to spend inordinate amounts of time trying to successfully matchmake students with appropriate mentors. This process would take a week, often more, making it incredibly tough for the students to ask for emergency assistance. With no strict set of criteria that can be consistently followed, a small team that is constantly juggling duties, and a great ambition, Greenwood is at a crossroads that may make or break them.
In order to keep up with their proposed expansion, it would only make sense to automate their current matchmaking process. They need something that will not only help them expand tomorrow but also a plan for 5 years from now.
That’s where we come in. We realize just how valuable a good matchmaking process is to Greenwood. The quality of a match can define the experience between the mentor and student. To ensure that, we have created a comprehensive plan that tackles both, Greenwood’s immediate and long-term needs.
Stage 1 will be our current algorithm, which will deal with the specific problems Greenwood has seen so far, which we will fine-tune upon working closely with them over the next couple of weeks/months. The next 2 stages are about implementing machine learning (simplistic followed by complex, as students and mentors increase), with a final hopeful far future plan wherein the AI may help Greenwood in improving other aspects of their internship program other than just the matchmaking. Our intention is to work with Greenwood and come up with metrics that may be of use to them to accurately judge how well a mentorship went (the most obvious of which would be to ask for some kind of feedback from students/mentors) to come up with data points that will give us insight into future operations. Our team sees this as an opportunity to grow with Greenwood and hopefully make a positive change in society.
All of us have been students in the US, though based in India, and we applaud Greenwood's initiative. Hopefully, in the future, we may be able to replicate such a program in our own country as well.
What it does
Greenwood specifically asked us just for a matchmaking algorithm, but we wanted to go the extra mile and give them a full technical plan. We propose a three-step plan that considers the size of Greenwood’s student count as it grows with time.
Stage 1: Current Algorithm
Dynamic personalizing Algorithm and basic UI (prototype available on Github) At this point, even though manual matchmaking is incredibly tedious, a machine learning model is not quite feasible because records of previous data are still being digitized. Until there is enough data, we’ll implement an algorithm that cross-checks every mentor profile with every student profile. Our algorithm can be implemented instantly by greenwood to aid their problem of matchmaking a large number of applicants. The algorithm is coded in python, easy to understand, and use language. The algorithm is sophisticated, but clean coded, making it extremely easy to read and understand for any programmer at Greenwood.
We have prototyped an algorithm and designed a UI sample. The algorithm uses three sources of inputs: Greenwood’s Database, a Mentor UI, and a Student UI.
The algorithm can be found in our Github repository. There exist CSV files to validate the python function.
Input 1: Greenwood Database Greenwood is storing/will store student info that they get from student applications in a database. From there, we are extracting students' technical skills, education level, and industry of preference.
Input 2: Mentor UI From an easy to use UI on the mentor landing page, we gauge their gender, technical skill, availability, the industry experience that they have, and preferred education level of their student. There are some more features we intend to work with if/when we work with Greenwood in the future such as their past with Greenwood as students/mentors, their current college year, the job they're in, etc. All these features will be thoroughly worked on when we're able to work closely with the members of Greenwood.
Input 3: Student UI
To take into account the comfort level a student can have with different types of people, our UI asks them a couple of simple questions, whose answers alter the algorithm to yield personalized results.
With these features in place, the algorithm gives the students an ordered list of mentors they can choose from. To ensure complete control, students also have filters if they feel the results do not accurately represent their needs.
Stage 2: Basic Machine Learning Model
The basic machine learning model for when greenwood reaches a mid-size client base As the number of students rises past 1000, the algorithm must adapt to keep up with the new dynamics of a growing company. At this point, greenwood would have just about enough data to train a basic machine learning model. This would be a recommendation algorithm that utilizes a simple machine learning technique used on low to medium amounts of data. This model will improve on the previous algorithm’s processing speed and matchmaking procedure to generate better matches for each student. Depending on the amount of features present within the data, different kinds of machine learning models can be utilized.
Stage 3: Complex AI
Advanced AI/machine learning model to handle extremely large datasets Once Greenwood has a massive reach, we can further adapt the recommendation algorithm to use complex technical AI. This AI will provide exceptionally fast processing speeds and approach the dataset as objectively as possible, giving students the matches they truly deserve. Again, students will be able to filter out personal preferences from the recommended list of profiles.
Step 3.5: Potential Impact
In the long term, the AI will no longer just help in matchmaking, but also find new opportunities such as unexplored geographical areas, underutilized tools, and most effective mentorship methods. This could springboard a bunch of other methods to further Greenwood’s goal of increased representation of Black and Latinx communities in the business world.
How we built it
The algorithm was constructed and programmed with the utmost care, taking into consideration a list of criteria provided by Greenwood that they consider vital for creating good matches. Each criterion was then weighted appropriately within the algorithm to give certain criteria a higher preference for generating a good match within the algorithm. Applicants rate the importance of each criterion on the preference UI page we designed, the algorithm then tries to match the applicant's preference ratings with the appropriate mentor profile. This generates a percentage match for each applicant with every mentor, thus achieving our goal, generating recommendations that are personalized to each applicant's needs.
Applicants can then view every single mentor present within Greenwood's database, with their best matches displayed in descending order. Applicants also have the option to filter out profiles based on criteria like gender, ethnicity, etc to help them get mentors they may feel comfortable with/are a better match.
Challenges we ran into
We found it challenging to try and balance the algorithm. Adding weights to each criterion was a complex task. Too high a weight gave too much preference to a criterion and visa versa. Moreover, trying to decide which criteria were more important than the other took quite a bit of time, however, after having had several discussions with the Greenwood staff, we were able to come to a conclusion.
Another important challenge we ran into was the issue of security. However, Greenwood specifically mentioned that they are working on developing an app themselves and all security measures, including authentication, will be implemented by them as they develop their application.
Unfortunately, Greenwood did not have any prior data records, therefore, pseudo-CSV data had to be generated to validate the algorithm's functionality. Moreover, their web app, which is currently in development, was also inaccessible. This meant we could not commit to creating a fully functional UI for their web app, given the chances that our UI would not fit right with their app, either in terms of functionality or design. Greenwood only wanted us to design a working Algorithm which we successfully achieved.
However, we still developed a sample UI with no core functionality. The purpose behind it is to show Greenwood how our algorithm could potentially be fully utilized in the event they need help with placing it in their app.
Accomplishments that we're proud of
The conclusion that we drew in the previous section is essentially the algorithm we designed. We are proud of how it functions. Albeit not perfect, it does a decent job for having only worked on it for over a week. We are sure we will be able to further optimize the algorithm if we continue to work with Greenwood.
What we learned
The unique dynamic of trying to empower Black and LatinX communities, introduced to us by Greenwood, taught us many invaluable lessons on problems faced by these communities. It was a pleasure to work with Greenwood to try and devise an algorithm that can potentially help these communities gain the mentorship they need to kick start their careers. Careers they might never have had the option of pursuing it was not for Greenwood.
What's next for Greenwood Project - Opportunity Hack
If we win, the financial funding will be enough for us to dedicate our time to the Greenwood project. Along with improving the UI experience, we can further optimize the algorithm to generate better matches. Better the match, better the mentor-student experience, and thus better the rep for Greenwood.
Once Greenwood collects and digitizes enough data over the next few months, we can test out multiple machine learning models, providing better matches at a faster pace.
At the end of the day, we want to make sure that the algorithm and model help accelerate Greenwood's growth and reputation. By trying our level best to ensure the best possible mentor-student matches, we want to stimulate extremely positive experiences for Greenwood's clients. We hope that word of these positive experiences spread around, leading to a greater number of applicants for Greenwood.