In my research work, I found that one of the most tedious yet essential skills to develop in the lab is conducting literature reviews to understand the current state of knowledge on a topic. To streamline this process, I've been building and refining Kanopik - an AI assistant that gathers and summarizes the most relevant papers to any scientific question. It pulls sources from trusted sources like PubMed and Semantic Scholar, filters them for relevance, and generates clear summaries and a paper list to help anyone get up to speed on a selected topic.

This code generates natural language captions for images using a custom deep learning pipeline built in PyTorch. A pretrained ResNet18 model encodes each image into a 512-dimensional feature vector, which is passed to an LSTM decoder that generates captions one word at a time. The model is trained on the Flickr8k dataset using padded sequences and optimized with cross-entropy loss. Inference is done via greedy decoding.

Vibe Maker is a web app that generates Spotify playlists from natural language prompts using OpenAI’s GPT-4o and the Spotify API. Built with Streamlit, it lets users log into their Spotify accounts and type in abstract or specific requests like “give me something upbeat to cook to” or “I want to listen to 90s rock.” The app converts these into structured Spotify search queries (e.g., genre:lo-fi year:2020-2025), fetches track results via Spotipy, and creates a new playlist directly in the user’s account.