AI Engineer · Independent research · Medical imaging & Autonomous driving
guillaume-cassez.fr · ORCID · Zenodo (5 papers) · OpenAlex · 💻 Weights · LinkedIn
AI Engineer based in Amiens, France, with a domain-aware and application-driven approach to deep learning. I focus on practical performance over headline benchmarks, and on clinically meaningful metrics over aggregate scores.
My research habit is the hard one: evaluate under the official challenge metrics (BraTS-2023 lesion-wise Dice/HD95, Cityscapes official mIoU), on full 5-fold cross-validation, with a pre-specified primary endpoint, Holm correction, effect sizes and bootstrap CIs — and publish the null results in full. Two of the five papers below report a loss that does not beat its baseline alone; what they establish is where it sits and what it is worth in combination.
- Medical imaging — 3D brain tumour segmentation (BraTS 2023 GLI, MedNeXt-B / nnU-Net v2, 1 196 patients): auxiliary losses (signed-distance transform, Kervadec boundary, blob), expert-initialised mixture-of-experts, parameter-free connected-component consensus.
- Autonomous driving — full-resolution Cityscapes segmentation (ConvNeXt-V2 + UPerNet): controlled 2×2 loss ablations, 4 configurations × 3 seeds, 160 epochs.
- Production ML — customer call transcription pipeline with
faster-whisper(local GPU inference). - Founder of ImmoIA, a PropTech AI product.
| Paper | Task · scale | Headline result (official metrics) | Artefacts |
|---|---|---|---|
| 1 · Two Models That Agree Beat the Best of Them Alone (v10, 2026-09-22) | BraTS 2023 GLI · 5-fold CV, n = 1 196 | A parameter-free connected-component consensus is the only configuration that beats the baseline on both official ranking metrics: lesion-wise Dice +0.024 (Holm p = 4.5×10⁻¹⁶), lesion-wise HD95 −9.49 mm (Holm p = 5.7×10⁻²⁶), ~41 % fewer spurious lesions at negligible recall cost | DOI · code · page · weights baseline / distmap |
| 2 · Precision Pays (v2.6, 2026-09-22) | BraTS 2023 GLI · 5-fold CV, n = 1 196 + 3 seeds | Fixed-weight boundary loss (λ = 0.05) is null alone (lesion-wise Dice +0.000, Holm p = 0.34) but is the precision mirror of the SDT head — and wins as a consensus veto: lesion-wise HD95 −9.95 mm (p = 2.3×10⁻²⁰) vetoing the SDT model, lesion-wise Dice WT +3.89 pp (p = 0.005) as primary model with a baseline veto | DOI · page · PDF · weights |
| 3 · The Gate Does Not Choose (v0.13, 2026-09-22) | BraTS 2023 GLI · 29 arms × 5 folds, 239 patients/fold | An expert-initialised patch-wise MoE is the best arm of the study: +0.0057 lesion-wise Dice over its own baseline averaged over 5 folds, positive on 5/5, above the pre-declared smallest effect of interest (0.003) on 4/5; the best of 24 training-free consensus operators reaches only +0.0013 (gate margin +0.0044 = 1.5× SESOI) — yet routing stays near-uniform, so the gate does not choose | DOI · page · PDF · weights |
| 4 · Boundary Loss Ablation for Full-Resolution Cityscapes (v0.2.2, 2026-06-28) | Cityscapes 1024×2048 · 4 configs × 3 seeds, 160 ep | CE+Boundary reaches the best mean mIoU 81.69 ± 0.25 (Boundary F1 77.32 ± 0.13) and beats the field-standard CE+Dice under Holm (p = 0.032) on a paired image-bootstrap over the 500 val images | DOI · code · page · weights |
| 5 · Distance-Map Auxiliary Regression for Full-Resolution Cityscapes (v0.1.0, 2026-06-28) | Cityscapes 1024×2048 · 4 configs × 3 seeds, 160 ep | CE+DistMap (SDT auxiliary head, dropped at inference → zero test-time cost) reaches the best mean mIoU 81.64 ± 0.27, beating CE+Dice (Holm p = 0.046) and the joint variant (p = 0.001) | DOI · code · page · weights |
Concept DOIs are cited above: they always resolve to the latest version of each deposit. Co-author on the three BraTS papers: Stanislas Larnier (methodological guidance and reviews).
- 🧠 3D BraTS viewer — 1 196 patients, 4 experts, 12 consensus operators and the patch-wise MoE, side by side with ground truth (webgl, no install).
- 📊 Per-patient ranking — Dice and HD95 for every one of the 1 196 validation cases, sortable and filterable.
- 🚗 Cityscapes viewer — full-resolution predictions explorer.
- 💻 Model weights on Hugging Face — 5-fold MedNeXt-B checkpoints (BraTS: baseline, DistMap, Kervadec boundary-loss, MoE-V3) and 3-seed ConvNeXt-V2+UPerNet checkpoints (Cityscapes),
safetensors, MIT.
PyTorch · nnU-Net v2 · MedNeXt · mmsegmentation · faster-whisper · Three.js · Next.js · TypeScript · Docker · Docker Swarm · FastAPI · AWS S3/CloudFront
Workstation Ryzen 9950X3D + RTX PRO 6000 Blackwell 96 GB + RTX 3090. Local-first ML training and inference.
MedTech R&D positions in France (AZmed, Gleamer, Pixyl, Owkin, Therapixel, Hera-MI, Milvue, Incepto, Raidium, etc.).