CraveAI (Prototype)

CraveAI is a project that leverages machine learning to provide personalized restaurant recommendations. By analyzing user's location and preferences, it suggests relevant dishes and restaurants.

Motivation and outcome: The goal was to create a system that understands user interests deeply and curates content accordingly, making it easier for users to discover new and engaging places to eat.

MAL Anime Score Predictions

This project pairs a data pipeline and model stack written in Python with a small React front end. The backend pulls and cleans MyAnimeList data, turns synopses into TF‑IDF vectors (with dimensionality control) and combines those with engineered metadata features; models are trained in scikit‑learn and with XGBoost/LightGBM and served through a FastAPI endpoint on Uvicorn. The frontend (Vite + React + Tailwind) shows an anime card grid with cover art, MAL IDs, color‑coded score badges, and a CSV export for further analysis.

Motivation and outcome: as an anime fan I built this to improve my chances in MAL’s seasonal score prediction competitions — the project focused on practical gains (robust preprocessing, leakage‑safe validation, and repeatable experiments) rather than academic novelty.

Approach and reasoning: I found that combining textual signals from TF‑IDF (reduced to control dimensionality) with structured features like genre, popularity, and airing metadata produced the most reliable signals. To get realistic performance estimates I used time‑aware / leakage‑safe splits, extensive error analysis, and K‑fold cross validation with early stopping during hyperparameter searches.

Effort and iterations: the work took several weeks and dozens of hours of training and tuning — many experiments were short runs to validate feature ideas, while final models required longer training and careful hyperparameter sweeps. I used feature‑importance and targeted error analysis to guide each iteration and selected the final model by RMSE/MAE on held‑out data.

MAL Anime Score Predictions

This project pairs a data pipeline and model stack written in Python with a small React front end. The backend pulls and cleans MyAnimeList data, turns synopses into TF‑IDF vectors (with dimensionality control) and combines those with engineered metadata features; models are trained in scikit‑learn and with XGBoost/LightGBM and served through a FastAPI endpoint on Uvicorn. The frontend (Vite + React + Tailwind) shows an anime card grid with cover art, MAL IDs, color‑coded score badges, and a CSV export for further analysis.

Motivation and outcome: as an anime fan I built this to improve my chances in MAL’s seasonal score prediction competitions — the project focused on practical gains (robust preprocessing, leakage‑safe validation, and repeatable experiments) rather than academic novelty.

Approach and reasoning: I found that combining textual signals from TF‑IDF (reduced to control dimensionality) with structured features like genre, popularity, and airing metadata produced the most reliable signals. To get realistic performance estimates I used time‑aware / leakage‑safe splits, extensive error analysis, and K‑fold cross validation with early stopping during hyperparameter searches.

Effort and iterations: the work took several weeks and dozens of hours of training and tuning — many experiments were short runs to validate feature ideas, while final models required longer training and careful hyperparameter sweeps. I used feature‑importance and targeted error analysis to guide each iteration and selected the final model by RMSE/MAE on held‑out data.

XY-Ball-Fight

A small browser-based 2D "ball fight" game built with plain HTML, CSS and JavaScript. This project demonstrates a simple game loop, basic physics, AI opponents, weapons, and an on-screen UI.

• Lightweight vanilla JavaScript implementation
• Simple physics-based movement and collisions (see engine/physics.js)
• AI opponents (see engine/ai.js)
• Modular engine files: game loop, weapons, UI, physics
• Easy to run locally — no build step required

Portfolio Website

My personal portfolio site built with plain HTML, CSS and JavaScript—showcasing my bio, projects, and resume link.

Chess-Bot

A Python implementation of a chess-playing agent using minimax with alpha-beta pruning and heuristics for board evaluation. (WIP)

Evodle

An idle‐clicker browser game powered by evolutionary algorithms—watch your creatures evolve and unlock new upgrades. (WIP)

NLP Analysis Tool

A Python package leveraging NLTK & SpaCy to perform text classification, keyword extraction, and sentiment analysis on unstructured data. (WIP)

VanklCommApp

A communication platform built with React Native and Firebase—features real-time chat, group channels, and file sharing.

Fitness-Run-Application-ACMERun

An Android app that tracks your runs, displays pace/heart-rate graphs, and syncs with Google Fit using Kotlin and Jetpack Compose.

Island Generator

A procedurally generated island terrain tool implemented in Java using Perlin noise and Voronoi-based erosion simulation.

Mesh Terrain Generator

A Java library for generating 3D mesh terrains from heightmaps, complete with UV mapping and customizable shader support.

Piraten-Karpen

A Unity-based 2D pirate adventure game featuring procedurally generated maps, inventory management, and NES-style pixel art.