Munich, Germany

I build machine-learning, forecasting, and control systems from the ground up. My projects are reproducible experiments, with efficient use of limited compute, calibrated uncertainty, and with a documentation of negative results as clearly as the successful ones.

A from-scratch 3-DOF simulation of powered-descent guidance for a reusable booster. It uses convex optimization both as an open-loop planner and as a closed-loop model predictive controller, then connects the landing burn to a full return profile based on public Falcon 9 telemetry.

• G-FOLD-style lossless convexification with successive convexification for aerodynamic drag
• Open-loop and MPC guidance compared under wind and state disturbances
• Atmosphere and drag model validated against public CRS-11 and CRS-12 telemetry
• Self-contained interactive 3D web visualization of the simulated return

A 97-million-parameter GPT written directly in PyTorch. The architecture, tokenizer, training loop, and evaluation suite are implemented from first principles, and the model was trained from random initialization on public text.

• Modern architecture with RoPE, SwiGLU, RMSNorm, QK-norm, and an independent implementation of the Muon optimizer
• Controlled comparison between two full training runs that differ only in optimizer and normalization
• Evaluation on ARC-Easy, ARC-Challenge, HellaSwag, and LAMBADA using a reimplemented standard log-likelihood protocol
• Results include the failures and data issues discovered along the way

A rigorous investigation of fine-tuning the Kronos financial foundation model on ETF data, including leakage-free calibration, walk-forward testing, and follow-up research after the model showed no directional edge.

• Split-conformal prediction intervals and statistical comparison against naive, GARCH, and EWMA baselines
• Negative directional result documented instead of optimized away
• Modest volatility-forecasting skill and independent portfolio experiments with trend following and volatility targeting
• Reproducible research pipeline plus a local forecast web interface

Predictive maintenance for aircraft engines using NASA's realistic N-CMAPSS DS02 flight profiles. The project compares classical machine learning and deep sequence models for remaining-useful-life prediction, with uncertainty rather than point estimates alone.

• Cycle-level feature engineering and sequence modeling with leakage-safe unit-based splits
• Gradient boosting and NGBoost compared with CNN/LSTM approaches
• Quantile, ensemble, and MC-dropout uncertainty estimates
• Evaluation with RMSE, the asymmetric NASA PHM score, interval coverage, and calibration

A coding-agent pipeline built from Llama 3.2 3B Instruct. It combines supervised reasoning training with an inference-time ReAct loop that executes generated Python, observes test failures, and retries.

• Reasoning SFT and general-assistant replay are complete
• GRPO uses unit-test outcomes as an objective, verifiable reward
• Sandboxed execution is shared by training rewards and the ReAct agent
• GRPO and the final before/after evaluation are not yet complete, so no performance gain is claimed