Our Story
As university freshers, we face the everyday problem of our daunting future. While people around us get interviews, others are rejected without being told why, most often after HireVue. We decided that enough was enough, and that a much-needed tool, to practice this craft of mastering interviews, needed building.
What it does
A Web App was the natural choice, seeing as the vast majority own or have access to a web-enabled device. This application allows users to create their own questions, specifying a preparation time and a recording time, to mimic the HireVue behaviour. Once the recording time has finished, the video is analysed by our backend and a start chart generated, highlighting the areas of strength and improvement. This is accompanied by a text explanation of how they got the scores for each category, along with improvements.
How we built it
Our system leverages a CNN to identify key body language cues, with samples taken periodically throughout the recording. We also utilise LLMs to analyse the transcript of the video, and providing feedback too.
Challenges we ran into
As we began developing our machine learning models, we faced issues of computational power for both training and inference, but particularly the former. The available datasets were minimal too. This could be a likely source of bias. We tried to mitigate this by diversifying the dataset with our own images and data augmentation techniques. This was also our first Web App development experience, so the team was unfamiliar with the work pipeline and development processes, leading to a steep learning curve.
Accomplishments that we're proud of
Being our first Hackathon, we are proud to have finished the project on time, while making the most of the social activities on offer too!
What we learned
We learned how to make Web Applications, integrating and developing AI models, working with databases and finally, the ability to work effectively as a team.
What's next for Virtually Hired
The most important first step would be to improve accuracy, through the use of larger and more diverse datasets, trained over longer training runs. In the longer term, it would be ambitious to try to expand the platform through supporting multi-users, tracking performance over time, and suggesting long-term strategies to improve.
Log in or sign up for Devpost to join the conversation.