Building

Overview Walter is an AI-powered trading agent designed to navigate cryptocurrency markets on Hyperliquid. It aggregates market data—including mid prices, candles, volume, funding, and trade pressure—into concise snapshots. Using an LLM backend (configurable with models like Gemini Flash or DeepSeek R1), Walter evaluates market context and enforces a strict confidence threshold before executing trades. It features: Autonomous Execution: Places market or limit orders via the Hyperliquid SDK. Safety Mechanisms: Configurable confidence thresholds and risk management parameters. Loop Architecture: Runs as a continuous scheduler loop for real-time monitoring. Key Features Market Snapshots: Aggregates comprehensive market data. LLM Decision Engine: Uses OpenRouter to access various LLMs for decision making. Risk Controls: Automatic leverage updates and order sizing. Python Native: Built with a modular Python architecture for easy extension. View on GitHub ...

Overview This project was developed as part of my Master’s thesis at the American University of Beirut (AUB). The goal was to characterize the construction period of buildings in Beirut using nothing but street-view imagery and Deep Learning. The system automates the extraction of building features to aid in socioeconomic analysis and disaster risk assessment. Technical Details Architecture: Fine-tuned Swin Transformer (SwinT) model. Pipeline: Includes data preprocessing, transfer learning optimization, and Bayesian hyperparameter tuning. Interpretability: Integrated LIME and SHAP for model explainability. Deployment: Dockerized application with a Gradio interface, also deployed on Hugging Face Spaces. View on GitHub ...

Overview DAURA (Detection of Disadvantaged Urban Areas) is a deep learning framework designed to identify vulnerable urban regions using satellite imagery. It orchestrates an end-to-end pipeline from patch extraction to model training and inference. Capabilities Patch Extraction: Automated sliding-window extraction and dataset balancing. Model: ResNet50 classifier with configurable hyperparameters. Interactive UI: Gradio app for real-time inference heatmaps. Integration: Daura operates as an installable python package for modular use. View on GitHub ...

Overview Investigating the dynamics of financial bubbles through data analysis and simulation. (Link pending update) View on GitHub

Overview Chrono-Brush is a Python library I built to handle the dirty work of time-series analysis. It provides scikit-learn compatible transformers to build clean, reproducible preprocessing pipelines. Features Outlier Removal: Statistical methods (IQR, Z-score) to clean noise. Imputation: Strategies for handling missing values (Forward fill, etc.). Trend Removal: Prophet-based trend decomposition. Composable: Fully compatible with sklearn.pipeline.Pipeline. from chrono_brush import OutlierRemover, MissingValueImputer, Pipeline pipe = Pipeline([ ("impute", MissingValueImputer(strategy="ffill")), ("outliers", OutlierRemover(method="iqr")), ]) cleaned = pipe.fit_transform(series) View on GitHub ...