An open-source local tool for exploring embedding datasets: embed → UMAP → cluster → label →
scope. A WebGL scatterplot map is the hero surface, surrounded by dense panels (cluster
labels, filters, data tables), a multi-step Setup pipeline, Jobs monitoring, and Compare
views. Python backend (Flask/waitress, ls-serve), React + Vite frontend in web/.
ML/data practitioners running long analysis sessions on their own machines. They are fluent in tools like Observable, Linear, Jupyter — they trust precise, quiet, dense interfaces and distrust decoration.
Product — design serves the task. The data is the star; chrome recedes and frames it like a scientific instrument.
Web (desktop-first, with a mobile Explore variant). Light and dark via
prefers-color-scheme; dark is where the tool lives most.
"Amber Console" — see DESIGN.md. Terminal-native futurism: warm graphite surfaces, one amber phosphor accent (the copper brand re-tuned per mode), IBM Plex Sans for language, IBM Plex Mono for everything the machine measured, status-diode light language, three restrained sci-fi signatures. Coherence over novelty; every affordance standard.