This Streamlit chat app showcases a ReAct agent powered by GPT-4o function calling, giving the model step-by-step reasoning and tool execution in one interface.

The agent selects from Google Search, Wikipedia, Calculator, and DateTime tools to fetch live facts, solve equations, and handle date math before crafting the final answer.

This is a self-contained repo with modular tool registry, tests, and conda env: clone, run streamlit run src/app.py, or extend with new tools in minutes.

In this project, I deployed a random forest model and dashboard on AWS visualizing data and predictions for diabetes hospital readmissions.

In addition to interactive visualizations, the dashboard enables the user to upload their own data and download model predictions.

Of the various models trained and tested, random forest performed the best, and the two most important features predicting hospital readmission were the number of lab procedures and the number of medications for the patient.

In this project, we studied the impact of noisy samples and pruning neural networks on image and audio through the lens of the cognitive science model of graceful degradation.

My focus in the project was the audio data, for which I trained 1D convolutional neural networks to process raw audio and 2D ones to process spectrogram-transformed audio.

We found that our neural networks were quite resilient to pruning when retrained and could learn to adapt to noisy inputs.