Sticky Insights
A browser-native ML system that reads a wall of sticky notes and surfaces the themes hiding in them — embedding, clustering, and labelling them entirely on-device, with no note ever leaving the tab.
Embeds each note into semantic space (MiniLM-L6-v2), then runs a dual-algorithm cluster bake-off — k-means and agglomerative (UPGMA), winner picked by silhouette score. It scores keyphrases per cluster with a KeyBERT-style multi-signal method (centroid similarity × TF-IDF contrastiveness × coverage) and turns the top phrases into a short theme label via a small language model (LaMini-Flan-T5-248M). Three validation gates vet every label and fall back to the top keyphrase when one fails, so the output is always sensible.
The entire pipeline runs in WebAssembly on a single thread — no backend, no API keys, no telemetry, offline-capable after the first model download. It doesn't just sort notes; it names the themes and surfaces the shape of consensus — how many people raised each one — then hands the judgment call to a human, deliberately. Real on-device inference behind a privacy boundary that never breaks is the hard, uncommon part.
The line that matters isn't pipeline-vs-model — it's pipeline-vs-system. The moment a person reads a generated label and reacts, you have feedback that could tune the next run; closing that loop is what would make this adaptive rather than deterministic. That's the next experiment, not a thing that's done — and being precise about which half is shipped is part of the engineering.