I like building models that push boundaries, stress-test assumptions, and reveal what really drives performance. My work sits where:

• Generative models break symmetry,
• Representations become interpretable, and
• GPU constraints force creativity.

I enjoy turning complex ML systems into clean, efficient, and rigorous experiments. No fluff—just models, math, and good engineering.

• Exploring non-Gaussian noise (Simplex, rank-based Gaussianization) and how noise structure shifts denoising difficulty.
• Building anomaly scoring pipelines using reconstruction spectra + latent statistics.
• Designing structured latents (FiLM conditioning, spatial Z-maps, capacity annealing).

• Vision Transformers + GNNs for structured reasoning.
• Latent geometry analysis using t‑SNE, PPCA, KL trajectories, and internal feature probing.

🔹 Swapped Gaussian noise with Simplex noise, Gaussianized-by-rank to preserve structure.

🔹 Saw significant AUROC gains with identical architecture.

🔹 Developed diagnostics to verify no data leakage + no artifact-induced shortcuts.

🔹 Focus on why noise geometry changes anomaly separability.

🔹 Trained a Tuned Lens to study token-level representation flow.

🔹 Compared KL divergence trajectories to detect prompt injection patterns.

🔹 Tuned Lens consistently outperformed Logit Lens in decoding stability & injection detection.

🔹 Implemented a full EM loop for PPCA with missing values, fully vectorized for GPUs.

🔹 Benchmarked PCA vs PPCA vs mini-batch variants on massive synthetic datasets.

🔹 Fine‑tuned LLaVA with LoRA + quantization under strict memory budgets.

🔹 Evaluated through LLM judges (DeepSeek R1 / MedAlpaca).

🔹 Improved reasoning structure with deliberate prompting.

🔹 Designed a sliding-window factorization model for financial-market signals.

🔹 Integrated hyperparameter search to stabilize dictionary evolution.

🔹Applied the approach to financial time‑series forecasting, supported by rigorous preprocessing and benchmarking against baseline models.

Models: Diffusion, VAEs, Transformers, GNNs

Training: LoRA, quantization, pruning, EMA, DDP, gradient checkpointing

Math: ELBO, variational inference, spectral analysis, latent-variable modeling

HPC: Slurm, A100 GPUs, CUDA profiling, memory debugging

Frameworks: PyTorch, timm, MONAI, PyTorch-Geometric, vLLM, Flash-Attention

• Understanding how noise geometry reshapes generative-model learning.
• Pushing more efficient training pipelines for large models.
• Strengthening my research engineering profile for roles involving foundation models, interpretability, and generative modeling.

• LinkedIn: linkedin.com/in/omar-arbi

Thanks for stopping by! 🚀