I build machine-learning experiments and developer tools. My projects explore image generation, language-model training, and ways to make experiments and coding-agent sessions easier to inspect and reproduce.
| Project | Focus | Where to start |
|---|---|---|
| ImageLab | A single-GPU experimentation framework with staged validation, checkpoints, and a run ledger. | Examples |
| CVQ | An unofficial channel-wise vector quantization implementation with MNIST samples and ablation tests. | Results and validation scope |
| ImageGen / EOSTok-pytorch | Caption-conditioned image generation and experiments diagnosing tokenizer collapse. | Diagnosis and controlled probes |
| autoresearch-mlx experiments | A fork with Apple Silicon optimizer and architecture experiments, charts, and a research journal. | What this fork adds |
| Helm / pi-kit | A terminal interface for Pi workflows, background tasks, and session history. | Helm interface and controls |
| dsh-archify-live | Map successful agent edits onto architecture components and surface stale source mappings. | Design and installation |
- StepTronOSS experiments: SDFT and DiPO additions to an upstream language-model training framework.
- LM research template: JAX/Equinox training infrastructure adapted from TTT-E2E.
- Pokémon EqM: image-generation experiments built around Equilibrium Matching.
Research repositories distinguish demonstrated results from experimental configurations. Fork READMEs identify the upstream project and the changes added here. Archived repositories preserve earlier experiments and learning projects.