ShelfFinder proof story
A grocery surface shaped around intent, not search.
ShelfFinder is the second proof surface because it shows product instinct outside cafeteria operations. The work starts with a simple frustration: grocery search boxes are bad at how people actually shop. A person thinks in lists, substitutions, store layout, and route. ShelfFinder turns that messy intent into a navigable task.
Spoken or typed grocery intent can include several items at once, not one search at a time.
The app separates a natural list into individual shopping tasks that can be matched, refined, and routed.
The output is about where to go and what to do next, not just whether product text matched a keyword.
Browser Web Speech and Deepgram are both accepted voice paths; Jinnia proves reusable streaming STT, utterance-end buffering, endpointing, and Azure proxy patterns.
The product begins with how shoppers think in the aisle: lists, substitutes, categories, and where to walk next.
Voice matters only when it removes friction. ShelfFinder keeps the output practical: a route, a match, and the next shopping move.
The voice work can borrow proven patterns from Jinnia instead of rebuilding streaming STT behavior from scratch.
ShelfFinder shows the same workflow judgment outside hospital food service: find the task shape first, then build around it.
ShelfFinder proves voice-first product design, multi-item parsing, route-aware shopping, consumer workflow instinct, and reusable AI voice infrastructure without turning the homepage into a demo gallery.