AI student at Howest, Kortrijk. Anomaly detection and interpretability. Every number below is measured, reproducible from its repo, and re-checked weekly by CI — including the ones that make me look bad.
Open to AI and backend internships.
SpuriousAD — image AUROC stays 1.000 at every confound correlation; the faithfulness collapse was an artefact
GraphCiteFaith — 3,965 of 3,965 cited node ids real, while two models name the structure at chance
DriftHarm — harm-precision 0.512–0.593 against a 0.517 base rate, so it is not a ranking
RingFaith — explainer overlap correlates r=+0.945 with its own random baseline
LeakGraph — only 5 of 10 GNN cells resolve above noise; the metric mostly tracks graph size
AI Act fairness audit — Annex III exempts fraud detection, so the audit was never legally required
Forecast backtesting — the split costs 4.1% MASE where the horizon costs 42.1%
M4 prediction intervals — nothing reaches nominal 95% coverage; the best is 86.6%
Explainable defect detector — PatchCore reproduced across all 15 MVTec categories, 0.9874 mean AUROC · demo
EU AI Act RAG — 90.2% faithfulness over 45 questions, graded by a separate model family · demo
Fraud decision trail — 0.1044 AUC separates the flattering split from the defensible one · demo
MLOps pipeline — prediction PSI correlates −0.709 with the AUC loss it should predict
RL reward shaping — six reward functions, two the agent exploited; 43.2% against a 73.5% oracle · demo
Hallucination-aware captioning — BLIP hallucinates on 4.5% of probes but misses 29% of objects present · demo
A/B testing & causal inference — mSPRT proved more conservative than specified: 0.9% error against a nominal 5%
LoRA fine-tuning — an adapter of 0.28% of parameters still degraded held-out behaviour
Offline vs online metrics — an AUC of 0.5386 next to what the policy actually earns
ARC-AGI-2 — 3.9%, published because a negative result honestly reported is still a result
Smart IVC cage platform — full stack, firmware to frontend; 173 of 204 commits mine on a 3-person client project
Gold medal, International STEM Olympiad — mathematics, France 2022. Bronze, German, 2023. 0.9086 private leaderboard on IEEE-CIS, scored by Kaggle against labels I never saw. Taught ~200 students over four years. ORCID 0009-0001-8746-4582.
Python · PyTorch · LightGBM · Transformers · SHAP · FastAPI · Streamlit · React · Docker · GitHub Actions · Raspberry Pi