💡 Inspiration

We’ve all had that one brilliant professor who is impossible to understand. Whether it’s a heavy accent, rapid-fire speaking, or complex jargon, trying to keep up can feel like a losing battle. For non-native speakers and Hard-of-Hearing students, this isn't just annoying—it’s a barrier to education. Current tools like Zoom fail because they are "brainless"; they hear the sound "Eye-gun-values" but don't know the professor is actually teaching "Eigenvalues." We realized the problem isn't the audio—it's the context. We asked: "What if the AI could study the syllabus before class starts?"

🚀 What it does

LectureBridge is a live captioning tool that "learns" your course material instantly. Upload: Drag and drop your syllabus or slide deck (PDF). Learn: The AI scans the document to build a custom dictionary of technical terms (e.g., "Bernoulli", "Stochastic", "Mitochondria"). Listen: It captions the lecture in real-time, using that dictionary to fix errors before they appear on screen. Clarify: It highlights these technical terms in Gold, giving students visual confirmation that the notes are correct.

⚙️ How we built it

We used a "Context-First" architecture to make the AI smarter. The Brain (Backend): We used Python (FastAPI) to handle the logic. When you upload a PDF, we use spaCy (NLP) to hunt for unique proper nouns and technical jargon. The Ears (AI): We feed those keywords directly into Deepgram's Nova-2 model. This "biases" the AI, forcing it to prefer words from your syllabus over generic English words. Simple Logic: Audio("Mass") + Syllabus("Calculus") = "Math" The Face (Frontend): Built with React and Vite. We faced a huge challenge with text flickering, so we built a custom "Flicker-Fix" engine using Framer Motion. It displays "gray" predictive text for speed, then snaps to "black" confirmed text for accuracy.

🚧 Challenges we ran into

The "Re-render Hell": Real-time text updates 20 times a second. Initially, this froze our browser. We fixed it by switching to Zustand, which handles high-speed data without refreshing the whole page. Audio Formats: Safari and Chrome handle microphones differently. We had to build a custom audio slicer to ensure the backend received clean, identical data from any device.

🏅 Accomplishments that we're proud of

Zero-Lag UI: We achieved a "perceived latency" of nearly zero by showing interim results instantly. It actually works: Watching the AI correct "Navier Stocks" to "Navier-Stokes" in real-time felt like magic.

⏩ What's next for LectureBridge

Professor Mode: A dashboard telling the lecturer if they are speaking too fast. Auto-Notes: Generating a markdown summary of the lecture automatically after class.

Built With

• deepgram
• fastapi
• gemini
• python
• spacy
• vite