I'm a final-year BSc Data Science, Protection & Security student at Thomas More University Mechelen, graduating September 2026. My work sits at the intersection of analytics, business intelligence, and applied machine learning. The kind that translates directly into decisions.
Last year I interned at Agiliz, shipping Looker interactive dashboards and leads generation pipeline creation. Since then, I've been building on my own: agentic research tools, early-warning systems, and weekend experiments with LLMs.
With amazing control over English, I write clean SQL, opinionated DAX, and care more about reproducibility than dashboards that demo well.
# currently working on
● building BE-Meds EWS — belgium medicine shortage early-warning system
# recently shipped
✓ shipped PowerBi MCP — MCP server that lets you talk to a Power BI workspace in plain English
# reading / learning
→ learning Microsoft Copilot Studio · LangChain · Agentic A.I. · Model Context Protocol
EOFMulti-agent AI system that turns a single stock ticker into a full risk-intelligence report. Four agents — Research, Analyst, Judge, Report — plus a ChromaDB memory layer, coordinated by an orchestrator and deployed live on Google Cloud Run. Uses an XGBoost model with SHAP explainability for the risk score, and a self-evaluating Judge agent that scores quality and applies guardrails before the report is delivered.
End-to-end agentic pipeline: the Research agent pulls live price data, fundamentals, and news through yfinance and Alpha Vantage, then writes a structured brief with Gemini. The Analyst agent computes RSI, volatility, momentum, P/E, and debt features, runs them through an XGBoost classifier, and uses SHAP to explain which factors drive the score. The Judge agent then evaluates the other agents' output on factual consistency, confidence, and completeness — combined with rule-based guardrails that don't rely on the LLM. The Report agent assembles a styled HTML report with embedded charts. Memory (ChromaDB) caches past analyses for instant repeat runs. Containerised with Docker and deployed to Cloud Run with secrets in Secret Manager and structured logging to Cloud Logging.
Microsoft Copilot Studio agent that enables researchers, engineers, and students to query sustainable material alternatives using natural language. The agent is powered by a Python ML backend, real scientific data from the Materials Project database (150,000+ materials), and SHAP-based explainable AI.
End-to-end machine learning system built on 284,000+ real credit card transactions. Detects fraudulent activity across a severely imbalanced dataset (577:1 ratio) using a tuned XGBoost model with SHAP-based explainability — deployed as a live, interactive web application accessible to non-technical stakeholders.
A full pipeline from raw data to deployment — automated quality gates, feature engineering, Optuna hyperparameter tuning, and MLflow experiment tracking. The Streamlit app lets analysts explore the data, compare model results, and run live predictions with no setup required.
Investment and risk dashboard built for the European Stability Mechanism, the eurozone's rescue fund. Consolidates position, exposure, and performance data across multiple facilities into a single decision surface for Middle Office analysts.
Python pipeline ingests fund and counterparty data into a SQLite warehouse, which powers three Power BI dashboards themed in ESM's visual identity. Drill-through flows let analysts move from portfolio-level KPIs down to individual positions, with careful attention to DAX performance on large position tables.
MCP server that lets you talk to a Power BI workspace in plain English — ask Claude what reports exist, inspect dataset schemas, trace lineage, and generate DAX measures from a natural-language description. No SQL. No DAX knowledge required.
Implements all three MCP primitives: nine tools (list, get, search,
generate) that Claude calls autonomously; six resources exposed as
powerbi:// URIs for passive reads; and three pre-built
prompt templates for dataset exploration, DAX explanation, and report
audits. Ships with realistic mock ESM/finance data so it runs immediately
with no credentials — plug in an Azure AD app registration and it
switches to a live Power BI workspace automatically. DAX generation calls
the Anthropic API with the dataset schema in context; a mock fallback
handles the no-credit case gracefully.
First hands-on MCP server build — exposes a pandas-based data-cleaning pipeline as a tool that any MCP-compatible LLM (Claude, GPT, local models) can call to clean tabular data on demand. Built to learn the Model Context Protocol end-to-end, from server definition to tool invocation.
The server wraps a cleaning routine that drops duplicates, imputes missing values with median for numeric columns and mode for categorical, and returns a log of every transformation applied. Deliberately kept the cleaning logic simple to focus the project on the MCP layer: tool registration with @mcp.tool(), type-annotated inputs, and structured string outputs that LLMs can parse and reason about. A CLI agent wrapper provides a non-MCP fallback for local testing.
Belgium Medicine Shortage Early-Warning System. Forecasts stock-out risk per molecule using public FAGG feeds, wholesaler data, and seasonality signals.
Ensemble of a baseline Prophet model and an XGBoost model over engineered features (dispensing volume, lagged shortages, substitution graph). SHAP explanations surface which signals drive each alert. Deployed as a Streamlit app with a weekly ETL job.
Actively seeking junior data analyst, BI, and data science positions. Interested in opportunities that align with my technical background. Responses within 24 hours.