Projects

LangChain-based LLM agent to autonomously generate, compile, and debug Java plugins for Google's Tsunami Security Scanner. Designed a two-stage workflow combining code generation with compiler feedback for self-correction, enabling fully automated plugin development for OWASP Juice Shop vulnerabilities. In evaluation, the agent produced 16 working vulnerability detection plugins, all without manual code fixes.

Comparative study of state-of-the-art deep learning models for multivariate EEG signal forecasting on the Temple University EEG corpus, comparing Transformer-based architectures including Informer, Autoformer, PatchTST, TSMixer, and Time Series Transformer. Designed a preprocessing and benchmarking pipeline for standardized EEG recordings, including PREP-based cleaning, resampling, normalization, and 22-channel bipolar montage conversion.

Ongoing thesis on designing and evaluating a graph-based Android ransomware detection system that enriches function call graphs with LLM-derived semantic node representations. Implements a static analysis and ML pipeline spanning APK analysis, graph construction, semantic feature engineering from code artifacts, and GNN-based classification, with systematic comparison of direct code embeddings versus LLM-generated abstractions.

Geospatial Computer Vision pipeline for automated building detection from multispectral satellite imagery. Integrated Sentinel-2 and OpenStreetMap data, generated segmentation masks, prepared stratified training/validation datasets, and trained PyTorch-based CNN and U-Net models with hyperparameter tuning and augmentation.

A web application that converts messy speech recordings into clean, optimized prompts for LLM chatbots. Features browser-based recording, Faster-Whisper transcription, and AI-powered refinement with support for both local (Ollama) and cloud (OpenAI) processing.

Project to detect station stops of Berlin commuter trains using only phone sensor data that apps can access without special permissions (magnetometer, accelerometer, gravity, and gyroscope streams). The data is resampled and transformed into time-based features that are less sensitive to how the phone is oriented and used to train an ML model that predicts stop intervals for each ride. Uses HistGradientBoosting with 328 engineered features including lag and diff features. Achieved 0.722 Jaccard score and top leaderboard position.

Deep learning system for video background reconstruction under virtual background leakage artifacts (e.g. Zoom or Teams Calls with virtual backgrounds). Formulated the task as soft segmentation/regression, generated continuous similarity masks as training targets, and trained a 5-channel U-Net on RGB, edge, and DoG features with a custom loss combining MSE, Gradient Loss, and SSIM. Implemented the full training and inference pipeline in PyTorch Lightning.

A robust backend system for complaint management built with FastAPI. Features integration with AWS services including WISE for payments, S3 for file storage, and SES for email notifications.

A simple web application to track scores during the card game Cabo. Features an intuitive interface for managing multiple players and rounds, making it easy to keep track of the lowest scores in this strategic card game.

An interactive educational web game designed to help players recognize psychological manipulation techniques. Through engaging storytelling scenarios, players learn to spot gaslighting, guilt-tripping, logical fallacies, and more by selecting the correct "Manipulation Cards" for each scene.

A local-first academic quiz workspace for self-paced exam prep. Organizes course-specific quiz sets, tracks progress directly in the browser, and needs no database or user login. The app can be filled via an Agent Skill that extracts lecture slides and converts them into verifiable quizzes, and supports the creation of personalized question formats.