Business & Data Analyst — turning customer, pricing and behavioural data into decisions people can act on.

Business Administration graduate from AUEB (Accounting & Finance), with machine learning training through AUEB's AI Data Factory bootcamp and a year of client-facing commercial experience in Athens real estate.

That combination is the point: I read a business question before I open a notebook, and I say plainly what the data does and does not support.

📍 Athens, Greece 📫 giorgstefanidis@gmail.com · LinkedIn

Junior Business Analyst / Data Analyst roles where the job is to answer a business question with data. Strongest fit: banking and credit risk, customer & CRM analytics, pricing and real estate analytics — the domains my background and projects already cover.

Gradient boosting scored 0.646 ROC-AUC on 2,996 banking customers. Logistic regression scored 0.502. That gap was the clue: the cleaning step had imputed missing values with a target-grouped mean, writing the answer into 13% of the rows. Corrected, the score fell to 0.527 — these features do not predict churn. The honest answer, and the reasoning that found it. pandas · scikit-learn · target leakage · cross-validation

Predicting sale prices from 79 property features. Six models benchmarked against a dummy baseline with 5-fold CV; Gradient Boosting won at RMSLE 0.1239. Includes permutation importance, residual analysis showing exactly where the model becomes unreliable, and a written report. Built on a year of working in the property market. scikit-learn · xgboost · pipelines · RMSLE

Data cleaning and exploratory analysis in SQL across US income and global life expectancy datasets, plus an Excel-based US debt tracker. Window functions, CTEs, aggregation and joins on messy real-world data. SQL · Excel · data cleaning · EDA

Six actionable customer segments from age, income and spending score, with k chosen by silhouette rather than by eye. Two segments — 32% of the base — drive the highest spending; the largest tier is where upsell has room to move. scikit-learn · clustering · marketing analytics

Matrix factorisation with SGD written from scratch on 1M MovieLens ratings — test RMSE 0.911 on a temporal hold-out. Includes a cold-start design for users with no history, and a Bayesian volume correction that reversed the genre ranking completely. numpy · pandas · matrix factorisation · cold start

The same task twice: a BiLSTM with attention built from scratch (75.3%) and a fine-tuned DistilBERT (80.6%), on an identical split so the comparison holds. The gain is largest on the rarest class — exactly what transfer learning is supposed to buy. pytorch · transformers · NLP

_{Also: Document Clustering — TF-IDF vs Word2Vec · unsupervised topic grouping where the better representation reverses between datasets.}

Data & BI · SQL · Excel · Power BI · Tableau Python · pandas · NumPy · scikit-learn · XGBoost · PyTorch · Transformers Methods · regression · classification · clustering · recommender systems · NLP Practice · cross-validation · leakage detection · pipelines · baseline-first evaluation Communication · written analysis reports · matplotlib / seaborn

Baselines first. A dummy model and a linear model go in before anything complex. Without them a score is a number with nothing to compare it to — and in the churn project, they are what exposed a result that looked good and was not.

Preprocessing belongs in the pipeline. Imputers, scalers and encoders fitted outside the cross-validation split leak information and inflate scores.

Report what the data supports. Two of the projects here end in a negative result. Both say so and explain why, because a conclusion that will not survive contact with production is worth less than no conclusion at all.

AI Data Factory — ML & Data Analysis Bootcamp, AUEB · Oct 2025 – Apr 2026 BSc Business Administration (Accounting & Finance), AUEB Real Estate Consultant & Assistant Sales Manager, Golden Home Real Estate · 2024 – 2025

English — Proficient (Michigan ECPE)