Skip to content

PRODUCT.md

Latest commit

 

History

History
69 lines (55 loc) · 3.53 KB

PRODUCT.md

File metadata and controls

69 lines (55 loc) · 3.53 KB
Raw

Product

Register

product

Users

ML engineers and researchers running LoRA/QLoRA fine-tuning experiments on open-weight models (Llama / Qwen / Gemma) on a single GPU — or on CPU for quick, keyless iteration. Their context: they run many small experiments and need reproducible, rollback-able, shareable run artifacts (datasets, checkpoints, adapters, configs, metrics) without standing up a heavyweight MLOps stack. They want to see exactly what each run produced and where it lives.

Product Purpose

An MLOps experiment-tracking dashboard for local LoRA/QLoRA fine-tuning with Unsloth, where every artifact a run produces is versioned on Backblaze B2 under a run-keyed prefix. It ships a Runs domain (create, start/re-run, compare, delete), live training progress with a loss curve, a scoped artifact explorer with presigned downloads, and a dashboard whose headline metric is the accumulating-artifact curve — B2 usage growing run over run. Fine-tuning is fully local (Unsloth on a CUDA GPU; a real, tiny transformers + PEFT fine-tune on CPU), so a run costs $0 and needs only B2 credentials. Success = an engineer can ingest a dataset, launch a run, watch checkpoints land on B2, and re-download any versioned artifact — reproducibly, without a database.

Brand Personality

Confident, precise, quietly professional. Voice is direct and free of hype ("Stop wiring boilerplate and start building"). The interface should feel like a modern developer tool — considered, calm, trustworthy — not a marketing showpiece. It is a neutral foundation that others rebrand: the design carries craft through restraint, not through a strong opinionated identity of its own.

Anti-references

  • Generic AI/SaaS slop. No gradient text, hero-metric templates, identical icon-card grids, tracked uppercase eyebrows, or decorative glassmorphism. These are the exact 2026 AI tells this kit exists to help builders avoid.
  • Over-branded / loud. No heavy brand-color drenching, decorative motion, or flashy effects. It is scaffolding to be rebranded, not a hero page.
  • Toy / prototype feel. No missing states, inconsistent components, or placeholder polish. Must read as production-grade.
  • Enterprise-drab. No Bootstrap-era gray boxes or dense-but-lifeless admin-panel look. Considered, like modern dev tools (Linear, GitHub Primer, Stripe).

Design Principles

  • Practice what you preach. The kit itself must model the production quality it asks agents to produce. Slop here propagates into every project built on it.
  • Neutral foundation, easy to rebrand. Identity lives in tokens (globals.css) and one config file. Screens are built from the shared UI kit so a rebrand is a token swap, not a rewrite.
  • Earned familiarity over novelty. Use standard, trusted affordances (top bar + side nav, command palette, data tables). The tool disappears into the task.
  • Every state is designed. Default, hover, focus, active, disabled, loading (skeleton), empty (teaches the interface), and error (says what's wrong + offers retry) — never half-shipped.
  • Consistency is the feature. One button vocabulary, one form-control set, one icon style across every screen. Divergence is a bug.

Accessibility & Inclusion

Target WCAG 2.1 AA. Body text ≥ 4.5:1, large/bold text ≥ 3:1, visible focus indicators on every interactive element, full keyboard navigation, correct semantic landmarks and heading order, labelled form controls, and a prefers-reduced-motion alternative for every animation. Full light and dark theme parity.