Nazaal

You can run end-to-end ML research: split train/val/test data without leakage, track experiments with MLflow or wandb, port PyTorch code to JAX, debug surprising metrics, manage SLURM cluster jobs, search literature, verify citations, and prepare LaTeX manuscripts with Overleaf git sync and arXiv camera-ready formatting.

ML researchers, PhD students, and research engineers working on probabilistic, Bayesian, or causal machine learning. The manifest reflects an ELLIS PhD workflow at Aalto's PML group, targeting people running GPU cluster experiments, writing academic papers, and maintaining reproducible research codebases.

Before launching, you state a hypothesis, pre-register predictions, and size seeds and ablations. After runs, you audit configs, logs, and metrics, then assign improve/degrade/noise/broken verdicts. Dedicated skills catch silent failures like data/loss/eval mismatch, overfitting tiny data, and convention bugs.

Skills reference concrete stacks: Python with pyproject.toml and virtualenvs, JAX ecosystem (optax, flax, equinox, blackjax, distrax, gpjax), hydra-zen configs, SLURM clusters with submitit, git with worktrees, Emacs Org agenda, LaTeX/Overleaf, and local utilities like the pzi paper library and cementic semantic search.

An evidence-before-completion gate blocks claiming work is done without fresh command output. Test-first development uses red-green-refactor with public-interface behavior tests, code review targets agent-written defects like deleted numerical guards and wrong axes, and root-cause debugging requires hypotheses before fixes.