Poppy is a chat assistant built specifically for teachers. It opens to a personalized home — a greeting, the day's weather and schedule — and a row of one-tap actions for the work teachers actually do: plan a lesson, write to parents, build a quiz, adapt content to a reading level. Conversations persist in a clean two-panel layout and title themselves automatically.
The point was to make an assistant that fits a teacher's hands rather than a generic chat box. The build leans on voice — real-time dictation with a live waveform — so you can talk a lesson plan into existence in the minute between classes, and on task-shaped shortcuts that skip the blank-page problem entirely.
Role
Solo build — designed and vibe-coded end to end with AI pair-programming. Next.js 16 + React 19 front end, GPT-4o-mini, a Storybook-documented component library, Playwright and Chromatic for tests.
What shipped
- Personalized home with greeting, weather and schedule, and quick-action shortcuts for common teaching tasks.
- Voice input via the Web Speech API with a live canvas waveform, so dictation feels responsive rather than fire-and-forget.
- Streaming GPT-4o-mini responses with conversation history persisted in localStorage and auto-generated titles.
- Responsive two-panel layout — navigation plus chat history — that collapses based on route.
- Accessible components documented in Storybook with an a11y addon.
Selected decisions
- Led with voice and task shortcuts instead of an empty prompt — the home screen answers 'what can this do for me today' before you type a word.
- Auto-titled threads from the first message so the history stays scannable without manual naming.
- Built the UI as a documented component library (Logo, Avatar, ActionGrid, NavItems) in Storybook, making the design system the source of truth.
Walkthrough
A closer look
A home, not a blank box
The first screen greets the teacher by name, surfaces the day's context, and offers direct actions — plan a lesson, write to parents, create a quiz, adapt content. It removes the hardest part of any assistant: knowing what to ask for.
Talk, don't type
Dictation runs on the Web Speech API with a live waveform drawn on canvas, so speaking feels acknowledged in real time. It's designed for the realistic moment of use — hands busy, a few minutes between classes.
Conversations that keep themselves tidy
Responses stream in, threads persist to localStorage and name themselves from the first message, and the sidebar adapts to where you are — so coming back to last week's parent email takes one glance, not a search.
A documented kit underneath
Every piece — avatar, action grid, nav, logo — lives in Storybook with accessibility checks, so the interface stayed coherent and inspectable as the feature set grew.