Athena

Main Logo

Inspiration

The journey of professional growth is never a solo endeavor. Behind every successful professional, there is a mentor who has guided them. But finding the right mentor isn’t always easy—many people struggle to connect with someone who truly understands their goals, experiences, and aspirations. Athena was built with one simple mission:

Connecting Talent with Experience

We wanted to create a mentorship platform where knowledge flows seamlessly—where students and professionals can find meaningful connections based on shared backgrounds, career aspirations, and interests. While many mentorship platforms rely on random pairings or keyword-based matching, we saw an opportunity to leverage artificial intelligence to build a smarter, more effective matching system—one that understands mentees and mentors beyond just a job title or a single skill. This multidimensional approach aims to provide the best mentorship to all women in STEM.

What it does

Firstly, it collects user information through a structured form and then uses OpenAI’s embedding model (ADA) to process responses, convert them into vector representations, and calculate the best mentor-mentee match using cosine similarity. The system ranks potential mentors and provides users with the best matches based on their academic background, work experience, and career goals. After determining best matches for mentees and mentors, Athena also provides a chatting platform for mentee's and mentors to connect to each other at a personal level, with the goal of eventually being a means to face to face meetings.

Key Features:

AI-powered matching algorithm
Secure authentication using JWT
Real time interactions provided by the Chatting Portal
Scalable backend with Flask and MongoDB

How we built it

Tech Stack:

Frontend: React, TypeScript, MaterialUI
Backend: Flask, MongoDB and JWT
Matching algorithm utilizes OpenAI's ada-002 model and numpy for calculations.

Challenges we ran into

AI Implementation and Embedding Generation

We went through extensive trial and error to fine-tune the embedding generation process, ultimately selecting OpenAI’s ada model for its balance of efficiency and accuracy. In addition to choosing the right model, we also refined the weighting system—assigning different levels of importance to various aspects of mentors’ and mentees’ profiles. This allowed us to generate more accurate and representative embeddings, ensuring that our matching algorithm truly reflects each individual’s background and aspirations.

Setting up / Optimizing MongoDB

While we had prior experience using MongoDB in other projects, this was by far our most extensive use case. One of the biggest challenges we faced was maintaining consistency in variable names across different parts of the backend and frontend. At times, we had three to four different variable names referring to the same data, which inevitably led to confusion and inconsistencies in the information being pushed to and pulled from the database. This issue resulted in unnecessary debugging and time loss.

Authentication

This was our first experience working with JWT—and authentication in general—so we had to navigate a learning curve, particularly when it came to retrieving the user_id of the logged-in user. Understanding how JWT handles authentication tokens and securely extracting user information was a valuable learning experience that strengthened our backend development skills.

Accomplishments that we're proud of

The Matching Algorithm

From our tests, the matching algorithm has performed surprisingly well. The concept of transforming user responses—especially an open-ended question—into numerical vectors was an exciting challenge, and one we weren’t entirely sure would work at first. However, by leveraging our knowledge of mathematics and OpenAI’s ada model to generate these vectors, we were able to develop a matching algorithm that far exceeded our expectations.

Secure Connections

Since this was our first time implementing authentication, we initially worried that we wouldn’t be able to integrate everything we had planned. We mistakenly assumed that working with JWT would be time-consuming and complex. However, after successfully implementing it at scale in a relatively large project, we now feel confident in our ability to handle authentication, and this accomplishment is something we are truly proud of.

Real World Applications

We strongly believe that Athena has significant real-world applications. As we developed this project, we realized that a platform like Athena presents a perfect opportunity for individuals like our own mothers—seasoned professionals who have accumulated a wealth of experience in their respective industries but are now retired. Through Athena, they can share their knowledge and mentor those who are in the same position they were in 20 years ago, creating a bridge between experience and emerging talent.

What we learned

Deadling with large databases requires extensive planning to a degree that we had not anticipated before.
Planning ahead especially with variable names is crucial to avoid time consuming bugs in the program.
Embedding using semantics and cosine similarity is a great way to compare particular elements.