🎓 M.Eng. in Computer Science, Artificial Intelligence & Mathematical Modeling (Label Mod'Math) — Université de Technologie de Compiègne (UTC) (Graduated Sept. 2026)
📍 Paris, France · 🌐 Open to roles in France, Switzerland, Germany, Netherlands, UK & Remote
🔗 Interactive Portfolio · LinkedIn · 📫 joshua.nancey@etu.utc.fr

EN: Engineer bridging rigorous mathematical modeling (stochastic processes, mathematical optimization, finite-sample statistical risk bounds, differential geometry) and production-grade AI & software systems (high-resolution medical vision, grounded clinical LLMs under strict GPU budgets, real-time collaborative servers, and OS-level Linux/Android security architectures).

FR : Ingénieur diplômé de l'UTC Compiègne (Génie Informatique — Filière IAD & Label Mod'Math), à l'interface entre la modélisation mathématique rigoureuse (optimisation, processus stochastiques, garanties statistiques de risque en échantillon fini, géométrie différentielle 3D/4D) et l'ingénierie logicielle & IA de bout en bout (vision médicale haute résolution, LLMs cliniques sans hallucination sous contrainte GPU, architectures distribuées et sécurité système Linux/Android).

• tn10-medical-ai-showcase — Epiconcept R&D Internship (6 months)
  • Part 1 (High-Resolution Mammography & Guaranteed Risk): Fine-tuned the interpretable GMIC (ResNet-22 global saliency + ResNet-18 local patch attention) on 300+ GB of 16-bit mammograms (RSNA, 2944×1920), boosting AUC from 0.683 to 0.889 (95% CI: [0.858, 0.921]). Implemented distribution-free Selective Classification (SGR / Hoeffding bounds) simultaneously bounding false-negative (FNR ≤ 5%) and false-positive (FPR ≤ 5%) clinical risks.
  • Part 2 (CATH-LLM — European ECDC Infection Surveillance): Benchmarked 20 local LLM configurations (MedGemma-27B, MedGemma-4B, Gemma-2-27B, Qwen-2.5-32B, Mistral-Small-24B) on MIMIC-IV under a single NVIDIA A100 40 GB budget. Built a deterministic verbatim grounding verifier (evidence_is_grounded, F1 = 0.91) and demonstrated zero-shot cross-lingual transfer to Croatian clinical notes (F1 = 0.88 via NLLB-200 / EuroLLM-9B).
• data-capsule-deep-piste — Co-authored an 8-chapter Dockerized interactive training capsule transferring deep learning, uncertainty estimation (MC Dropout, Deep Ensembles), and selective risk control to public-health epidemiologists and data scientists.

• surfaces-implicites — BMBI Lab Research Project (CNRS UMR 7338, UTC)
  • Modeled 3D & 4D space-time deformable biological membranes (red blood cells & micro-capsules in shear flow) via continuous implicit level-sets $f_\theta(x,y,z,t) = 0$.
  • Compared polynomial regression (recovering the exact Kuchel quartic in 10 coefficients vs. 197,886 neural parameters), 3D Fourier features, and 4D SIREN / Positional Encoding MLPs with Eikonal regularization ($|\nabla f| = 1$), achieving IoU = 99.65% on non-affine capsule deformations.
  • Computed exact Gaussian ($K$) and Mean ($H$) curvatures via 2nd-order PyTorch autograd without mesh discretization, verifying the Gauss-Bonnet theorem ($\oint K,dA = 4\pi$) to 0.025% precision.

• Contextual Anomaly Detection Platform — Aubay Innovation Lab Internship (6 months)
  • Benchmarked PyTorch Tabular architectures (GANDALF, FT-Transformer, TabTransformer) coupled with Isolation Forest and Deep Isolation Forest (DIF) on mixed numerical/categorical datasets (F1 = 0.903, ROC-AUC = 0.985).
  • Engineered a full-stack Flask REST API + React web platform integrated with AWS S3 and reproducible Conda/Docker environments.
• Machine-learning-prediction — Statistical Learning & Clustering (SY09): Logistic regression with Wald statistical tests, Euclidean/Bayesian classifiers, and Gower-distance / Adaptive Mahalanobis K-Means on clinical cardiac data.
• ETL-project — Data Warehouse & Carbon Footprint Analytics (NF26): End-to-end Python ETL pipeline and dimensional modeling on the ADEME Base Carbone dataset.

• monlycee-relay — Zero-Trust Cloud Session Broker & Manifest V3 Browser Extension: Offloads Keycloak SSO authentication to a remote Linux cloud microservice (FastAPI + Playwright Headless Chromium with asyncio.Lock concurrency and < 50 ms RAM session cache) and injects authenticated cookies into Microsoft Edge via a Manifest V3 Service Worker without ever storing credentials on the client machine.
• jarvis-career-engine — Personal Second Brain (SQLite FTS5 BM25 + Knowledge Graph) & Multi-Judge Grounded Career Engine: Queries 51 European ATS boards in parallel and filters offers through a 2-pass, 3-LLM committee enforcing deterministic verbatim quote verification (evidence_is_grounded).
• custos & self-system — Multi-Platform OS Hardening: Native Android Device Owner (DevicePolicyManager) anti-circumvention suite in Kotlin (cryptographic time-locked policies, dual watchdog services) and Linux kernel/daemon enforcement (systemd watchdogs, mitmproxy TLS inspection, dm-verity read-only rootfs).
• Simulation-go & Seven-wonder-duel-in-C — Multi-Agent System Simulation in Go (IA04) and complete object-oriented implementation of 7 Wonders Duel in C++17 & Qt 6 (LO21) (MVC architecture, autonomous AI agents, UML design).