Education

           

Master's in Business Analytics & Data Science · IE University of Science and Technology · Final Term · Advanced AI

My projects don't follow a single theme: I've built a Bayesian model for predicting box office outcomes, a real-time crypto pipeline, a movie recommender trained on millions of ratings, an MLOps pipeline for house price prediction, an AI travel planner, and a 2D platformer game. I just tend to go deep on whatever I find interesting at the time.

Mixed-Frequency Forecasting of Energy CPI: MIDAS Models, LASSO Regularisation, and State-Space Extensions

Research capstone benchmarking MIDAS regression variants against ARIMAX, XGBoost, and LSTM for forecasting US consumer energy inflation from weekly crude oil prices. Introduces three novel model extensions (CLM-SS, LASSO-MIDAS, Kernel U-MIDAS), uses a three-layer expanding-window validation design, and identifies a 5–7 week oil-to-consumer-energy transmission window. Built in R with the midasr package.

Languages

ML / AI

Data Engineering

MLOps

Web Dev

Visualization

Other

• Hermes - AI European Travel Planner
  AI travel agent that builds personalised multi-city European itineraries. Scores 80 European cities using a proximity-aware scoring system (activity match, budget fit, seasonality, and geographic routing from your departure city), estimates transport across flights, trains, and buses, pulls live Wikipedia descriptions and Open-Meteo weather data, and generates a 12-month pricing calendar. Runs fully locally using Ollama (llama3.1:8b) with no paid API keys required. Built with Python, Streamlit, LangChain, and Plotly. 90 automated tests.

• The Blockbuster Formula
  Bayesian Network model predicting box office outcomes (Flop to Blockbuster) on ~2,900 films from TMDb (2000–2024). Full pipeline: API collection, feature engineering (actor prestige scoring, budget tiers, release windows), EDA, probabilistic inference, and a baseline ML comparison.
  Deployed: blockbuster-bayesian.streamlit.app

• Datathon 2025 – 1st Place
  Won a time-bound datathon analyzing aviation traffic and CO₂ emissions across European airspace using EUROCONTROL data (2010–2025). Built in Python with Tableau dashboards, identified trends between flight activity and environmental impact.

• Reverie. A GRU and NCF Powered Watch-Next Engine
  Reverie is a movie recommender that predicts what a viewer should watch next from their watch history. A GRU handles the sequential side: given the order of movies watched, it predicts the next one and re-ranks instantly when a pick is rejected. An NCF model handles the rating side: trained on roughly 9.9 million Letterboxd ratings, it predicts how much a specific person will like a specific movie, which lets it score modern films the GRU never saw and slot in brand-new users via embeddings. It ships as a full product, a layered FastAPI backend with JWT auth and a React, Vite, TypeScript and Tailwind frontend, with accounts, watchlists, friends, and a Blend feature that merges two people's picks. Evaluation follows a strict correctness contract (leave-one-out testing, confidence intervals, Wilcoxon significance tests) and a pytest suite covers auth and business logic.

• VC Startup Pitch Evaluator — AI-Powered Due Diligence
  Built a multi-agent GenAI system that analyses startup pitch decks end-to-end using the Gemini API, LangGraph, and a ReAct fact-checking agent that validates founder claims against live web sources. Outputs a structured investment memo with dimension scores (market, team, product, traction) and flags low-confidence decisions for human analyst review before a final recommendation is made.
  Deployed: vc-analyst.streamlit.app

• Budget Calculator: Exchange Edition
  Live Streamlit app for tracking spending across countries and currencies. Started as a personal tool during my winter 2024 exchange in Budapest, rebuilt in 2026 with multilingual support (EN/FR/ES), geographic spend maps, file import/export, and multi-currency conversion.
  Deployed: budget-calculator-tracker.streamlit.app

• Smart Residential Price Estimation -Production MLOps Pipeline
  End-to-end MLOps system that predicts Ames, Iowa house prices from structured property data using a LassoCV regression model (R² ≈ 0.80–0.89). Built as a production-grade pipeline, not a notebook, with modular preprocessing, feature engineering, and training stages orchestrated via main.py, experiment tracking and model registry through Weights & Biases, and a FastAPI service exposing /health and /predict endpoints. Includes Docker containerization, GitHub Actions CI/CD (test and build on PR, auto-deploy to Render on release), reproducible conda-lock environments, and 54 tests at 84% coverage. Built for a graduate MLOps course to demonstrate reproducibility, monitoring, and deployment practices beyond model accuracy alone.
  Deployed on Render: API docs

• Personal Finance Analytics Package
  Python package that categorizes bank CSV transactions using a regex rule engine and produces monthly budget summaries. Immutable dataclasses, CLI via argparse, pytest-tested, packaged with uv and published to TestPyPI. Backend library for a companion Streamlit app. Built as a Python Data Analysis II course project.
  TestPyPI: pocketwise-finance

• Crypto Streaming Analytics
  Real-time pipeline using Apache NiFi + MinIO + Python for ingesting and analyzing live Bitcoin and Ethereum price streams. NiFi flows pull data continuously and land it in MinIO (S3-compatible object storage), with Python handling downstream parsing and analysis. The whole stack was deployed and run on a virtual machine via Oracle VirtualBox to keep the pipeline isolated from the host environment.

• Tower of IE: The Wizard Climb. Browser-playbale 2D platformer built with Python and Pygame
  A 2D vertical platformer built with Python and Pygame, playable in-browser via WebAssembly (Pygbag). Climb a tower by conjuring platforms in real time, with physics, collision detection, camera scrolling, and a JSON-based leaderboard. Modular OOP architecture.
  Deployed on itch.io.

• Marvin – Reinforced AI Chess Bot
  AlphaZero-style chess engine in PyTorch: dual-head ResNet policy/value network trained via MCTS self-play RL, with optional supervised pretraining on grandmaster PGN games and a Flask web UI to play against it with live Elo tracking. Includes a pytest suite.

Music: EDM · Rock · Indie · Alternative · Jazz · Disco · 80s Pop · Huge Queen fan
Sports: Volleyball & Basketball
Also into photography and video editing