Training & Pipeline Tooling

Overview

A feral garden of utility scripts, automation tools, and experimental code that supports everything else in this portfolio. Things come and go here. Some scripts run for months and become load-bearing infrastructure. Others are useful for three days before a new technique makes them obsolete. That’s the point — this is the alchemy lab, not the library.

Only the experiments that prove themselves repeatable and genuinely useful graduate into formal documentation and long-term use. The rest get composted, their useful parts absorbed into whatever comes next.

What Lives Here

Training automation — Configuration generation, dataset validation, checkpoint monitoring, and batch orchestration for running multiple LoRA experiments sequentially. The scripts that keep extended training campaigns running efficiently while we sleep.

Pipeline utilities — VRAM management, stage handoff scripts, and output validation for multi-model video generation workflows. These handle the resource allocation puzzle of chaining large models on a single GPU — the kind of problem that changes shape every time a new model drops.

Data processing — Preprocessing, metadata extraction, and quality filtering scripts used during dataset curation. The unglamorous work that determines whether a training run produces something useful or waste.

Experiments — One-off scripts testing new ideas, benchmarking unfamiliar tools, probing the edges of what a new architecture can do. Most of these die young. The ones that survive become the next generation of training automation.

Design Philosophy

Written for utility over elegance. The code that lasts handles errors gracefully, logs extensively, and can be interrupted and resumed. The code that doesn’t last just needs to answer a question quickly and get out of the way. Both are valuable. Knowing which is which is the skill.