Levi Reece

Developed a deep learning model capable of decoding speech from silent video using 3D Convolutional Neural Networks (CNNs), with the goal of improving communication accessibility for the deaf and hard-of-hearing community.

Working alongside a team of researchers at UCLA COSMOS, I helped build and train a CNN that analyzes short video clips of lip movement to classify spoken words with no audio required. We sourced and preprocessed a labeled dataset from Kaggle, applying grayscale conversion and resizing before feeding frames into our model architecture, which included convolutional, pooling, and fully connected layers.

Over 40 training epochs, we used categorical cross-entropy loss and backpropagation to iteratively refine the model's parameters, monitoring both training and validation accuracy to ensure generalizability. By the final epoch, our model achieved 89.84% accuracy on the validation set — a result visualized through confusion matrices showing steady diagonal convergence as training progressed.

This project deepened my understanding of how AI can be designed with real-world accessibility impact in mind, and highlighted promising directions for future work, including diverse multilingual datasets, mouth-region extraction, and confidence scoring for predictions.

A peer-to-peer social betting app: bet on anything verifiable with friends, groups, strangers, or yourself. The name comes from the river Styx, which Greek gods swore unbreakable oaths on. The idea is to put money behind your word.

Side bets between friends usually fall apart at payout. Someone forgets, argues about the result, or never pays. Styx Bet locks the terms and the money before the outcome:

• Stakes are held in escrow up front, so nobody welches.
• Participants confirm the result. If they disagree, AI reviews the evidence and makes the call, and either side can appeal.
• If the outcome can't be proven, the pot goes to charity instead of anyone's pocket.

It supports 1-on-1 bets, group pools with custom payout splits, personal bets against yourself (like hitting the gym five days a week), and collateral for dares that pays out once the group confirms someone followed through. Creating a bet is a simple form (terms, stake, odds, payout split, deadline), which keeps it fast and unambiguous. We earn from services around the bet, never from who wins.

We built the first prototype in Google AI Studio during a 3-hour hackathon, using React, Firebase Authentication, Firestore, and the Gemini API, deployed on Cloud Run. Real-money betting between people is regulated state by state, so finding the right licensing and payments path is our first priority.

Co-built an AI-powered outreach tool for BestMate, a field service business app, with a team of four at the UC Berkeley LAUNCH Hackathon, where we won. The tool finds potential leads, scores which ones are worth contacting, and manages the outreach process from start to finish. When a lead replies, a "Needs You" feature pauses the automation, summarizes the response with a suggested next step, and hands the conversation back to a human. I built the content side, using AI to generate and automate marketing posts that support the outreach. Our guiding principle was to automate the work, not the relationship.