Epsilon

Epsilon Landing Page
Multilingual Office Hours
Concept Quiz
On-Demand Office Hours
Video-Generated Lecture

Inspiration

As teaching assistants for data structures and algorithms from the University of Pennsylvania, Tanay and I were motivated to improve our students' learning experiences through more intuitive and on-demand explanations. When we were CIS 1210 students, we would sit across from each other and go topic by topic drawing out algorithms, justifying POCs, and proving runtimes. Luckily, we were friends who not only spoke the same language but also had time to help each other out like this.

However, this is not the case for many students. For example, while they may have great friends, they might not have enough time to visit limited office hour (OH) times and my have too little experience with learning in English, disadvantaging them in and beyond lecture. To these students, we one day hope to present Epsilon.

What it does

Epsilon is an AI Teaching Assistant that will take away Tanay and mine's salaries. Based off of class lecture notes, it provides on-demand video lectures, live OH, and review quizzes just like a human TA would.

With a simple description of a students' issue, a video explanation will be generated to give students a better intuition against the confusion they are facing. If any concerns remain, live OH allow students to directly argue their case and find flaws in their reasoning. Once they have their ideas refined, students can finally explain their intuition to AI TA's and get feedback on their understanding.

How we built it

The audio-based theme behind Epsilon is spurred by 11ElevenLabs, a text to speech software that allows creators to seamlessly integrate audio into any AI agents they may create. Using OpenAI x 11ElevenLabs x Sync x Fal, we were able to create a text to speech to video pipeline that allows students to get content-based video lectures. Through 11ElevenLabs API, we created a feedback loop between students asking clarifying questions in OH and Epsilon asking students content comprehension questions in Concept Quizzes. The combination of these three tools made Epsilon a "hollistic" TA.

Challenges we ran into

Many. We ran into many challenges as a two person team on an ambitious mission. Our first, feature selection. What actions make a TA a TA? Reflecting on my experience as Head TA and Tanay's first sem, we decided that conveying conceptual topics in easy terms was the hallmark of a great TA. With a goal in mind, we began rapidly prototyping voice agents we could use to interact with students.

Hence, our second issue. In order to achieve video lectures, we needed to go from prompt to text to speech to video. After preparing our knowledge base for OpenAI's API, we needed to transform our generated text into audio. Here, 11ElevenLabs' text to speech agent allowed us to simply move past that step and focus on audio to video. This is when, we discovered the power of Fal, which hosts models such as sync-lipsync, that allowed us to input video and desired audio to create a human mouthing our words.

Accomplishments that we're proud of

As it was mine and Tanay's first hackathon in sophomore year, we are extremely proud to have learned so much about using 11ElevenLabs, Fal, and many more agentic AI tools.

What we learned

In the end, we learned about 30+ startups creating groundbreaking tools to simplify the agentic AI era.

What's next for Epsilon

Epsilon has many features to wrap up. Future solutions to consider involve homework help without giving away answers and math drawings while listening to video-generated lectures.