Pavvy: AI Education Tool for Lectures Videos
Developed a tool that can process lecture videos, extract transcripts with key-word highlighting, summarization, and automated quiz generation using ML. Finetuned a DistilBert model, used GPT4, and set up a model pipeline using Vertex AI and GCS.
Multimodal Income Prediction Model
Developed a model to predict income percentiles in NYC using satellite and street-level imagery. Utilized a pretrained Swin Transformer with a regression head and fine-tuned it for accurate predictions. Experimented with contrastive pretraining and other advanced training strategies to reach state-of-the-art performance.
Uncertainty Aware Policy Gradient Methods
Modified PPO using a bootstrapped ensemble of critics to gather a measure of the uncertainty in the value function. Using this I developed novel algorithms for dynamically adjusting lambda in Generalized Advantage Estimation to balance the bias-variance tradeoff, as well as exploring dynamic step sizes that are linearly and inversely proportional to the value variance. Managed to outperform baselines in terms of mean and standard deviation of sample efficiency