• Designed a hybrid financial forecasting architecture based on the hypothesis that market prices represent a superposition of three distinct components:
  Fundamental signal derived from financial news and macroeconomic context
  Technical/chartist signal extracted from historical price patterns
  Noise, mean-reverting deviations from the combined trend
  To isolate these components:
  Long Short-Term Memory (LSTM) networks were trained on historical price data to extract and forecast the chartist (technical) trend.
  A Retrieval-Augmented Generation (RAG) pipeline was implemented to incorporate fundamental context by dynamically retrieving and embedding real-time financial news.
  The residual between actual prices and the combined trend (LSTM news-informed context) was interpreted as market noise.
  A mean reversion strategy was then applied exclusively on this noise component, under the assumption that it represents short-term inefficiencies oscillating around a transient equilibrium.
  The strategy was evaluated through live simulation using Interactive Brokers (IBKR)Designed a hybrid financial forecasting architecture based on the hypothesis that market prices represent a superposition of three distinct components: Fundamental signal derived from financial news and macroeconomic context Technical/chartist signal extracted from historical price patterns Noise, mean-reverting deviations from the combined trend To isolate these components: Long Short-Term Memory (LSTM) networks were trained on historical price data to extract and forecast the chartist (technical) trend. A Retrieval-Augmented Generation (RAG) pipeline was implemented to incorporate fundamental context by dynamically retrieving and embedding real-time financial news. The residual between actual prices and the combined trend (LSTM news-informed context) was interpreted as market noise. A mean reversion strategy was then applied exclusively on this noise component, under the assumption that it represents short-term inefficiencies oscillating around a transient equilibrium. The strategy was evaluated through live simulation using Interactive Brokers (IBKR)

Developed an AI-powered Video Assistant Referee (VAR) for handball using a ZED 2i camera and YOLOv4 machine learning model.
Implemented real-time object detection (handball tracking) with Python, Bash, and JavaScript.
Presented the system to professional handball teams to enhance training efficiency and decision-making accuracy during matches.