Projects

Things I built, shipped and occasionally broke first.

10 projects
01
Full Stack

FootballIQ

AI powered football analytics platform built on Clean/Hexagonal Architecture. 11 module system with medallion data warehouse, XGBoost player valuation (FIFA 2026 data: 48 teams, 1,248 players), and RAG powered analyst grounded in SQL. 154 tests, 92% coverage.

#Python#FastAPI#PostgreSQL#pgvector#XGBoost#SHAP#Next.js#Docker#dbt#GitHubActions
02
Data Engineering

ImmoDash

German rental market intelligence platform, rebuilt every Monday from 11 open datasets. dbt medallion warehouse on DuckDB/PostgreSQL with 75 data tests covering 37 cities and 400 districts, LightGBM quantile forecasts with conformal ranges (1.74% error at 12 months, 62% below a no change baseline), a 25 endpoint FastAPI, and LLM city summaries fact checked against the data.

#Python#dbt#DuckDB#PostgreSQL#LightGBM#FastAPI#Groq#Next.js#GitHubActions#Vercel
03
Data Engineering

Career OS

Production MLOps system with 13 autonomous nodes. Ingests 680+ real job listings nightly via Prefect ETL, an ATS resume analyser with MLflow, a RAG career advisor powered by Qdrant and Groq LLMs, and 30/60/90 day skill demand forecasting.

#Python#FastAPI#Prefect#MLflow#Qdrant#Groq#GitHubActions#Vercel#Render#Next.js
04
Data Analytics

German Tech Salary Explorer

Interactive Plotly/Streamlit dashboard exploring 1,400 German tech salary records. Filters by role, city, experience level, and company size with real time chart updates.

#Python#Plotly#Streamlit#Pandas#DataVisualization
05
ML/AI

German Day-Ahead Electricity Price Forecaster

End to end ML forecasting pipeline for German electricity prices using SMARD market data (8,664 rows). LightGBM reduced forecast MAE by 62% over baseline (34.00 to 13.01 EUR/MWh). Served via a FastAPI REST endpoint with Docker containerization.

#Python#LightGBM#FastAPI#Streamlit#Docker#SMARDAPI#RESTAPI
06
ML/AI

White Stork Migration Analysis

Behavioral state classification on 54,613 GPS locations from 42 white storks using Hidden Markov Models. Joint project with Universität Bielefeld and the Max Planck Institute of Animal Behavior. Also applied conditional logit models for habitat selection.

#Python#R#HMM#ConditionalLogit#GPSData#StatisticalModeling
07
ML/AI

CNNs vs. Transformers Sentiment Analysis

Comparative study fine tuning BERT base (Hugging Face, PyTorch) against CNN architectures on the SST-2 dataset (67,349 labelled sentences). Achieved 90.14% accuracy with BERT, using progressive layer freezing, regularization, and systematic hyperparameter search.

#Python#PyTorch#BERT#HuggingFace#TensorFlow#NLP#SST-2
08
ML/AI

Cardiovascular Disease Risk Prediction

Binary classification on 308,854 CDC BRFSS health records. Achieved AUC-ROC 0.836 using XGBoost with class imbalance handling, feature selection, and model interpretability analysis.

#Python#XGBoost#Scikit-learn#Pandas#CDCBRFSS#Classification
09
ML/AI

Quora Duplicate Question Detection

Progressive feature engineering pipeline on 404,290 Quora question pairs. Improved accuracy by 6.1% (72.0% to 76.3%) and F1 score by 29% (0.519 to 0.671) over a bag of words baseline using 15+ token level and fuzzy string features. Benchmarked logistic regression, random forest, and XGBoost.

#Python#Scikit-learn#XGBoost#NLP#fuzzywuzzy#FeatureEngineering
10
Robotics

Voice-Enabled Human-Robot Interaction System

Real time voice interaction pipeline for human robot communication. Integrated Whisper ASR, RASA dialogue management, and TTS within a modular ROS2 architecture on Ubuntu Linux, using a timer driven MutuallyExclusiveCallbackGroup for zero race conditions.

#Python#ROS2#WhisperASR#RASA#TTS#Ubuntu#SpeechProcessing