Scene Words

Context-based English learning, fully offline. A hackathon MVP that runs across a web browser, a FastAPI backend, and an ESP32 "Cheap Yellow Display" — no LLM API key, no cloud, no network required.

Live demo at the Aberystwyth hackathon — the handheld CYD board renders the same Dashboard that the web app does, both backed by the local FastAPI service.

What it does

Type or paste an English sentence — for example "Hi, can I get a latte with milk to go? How much is the large size?" — and a rule-based NLP pipeline:

1. Detects the scene context (Coffee Shop / Doctor / King's Cross / Baker Street) with a confidence score
2. Extracts useful keywords and phrases from the text
3. Generates a one-line summary of what's happening in the scene
4. Saves anything you tap into a SQLite-backed review library
5. Renders the dashboard on both the browser and the embedded CYD touch screen

Everything runs on local Python standard library + SQLite. No LLM, no API key, no internet.

Demo

1. Input — pick a source, type a sentence, or use a demo button

2. Analysis — context detected, keywords + phrases + summary in one response

3. Dashboard — saved items aggregated, each tagged by source context

Tech stack

Layer	Tech
Backend / NLP	Python 3, FastAPI, Uvicorn, Pydantic v2, SQLite
Web	Vanilla HTML / CSS / JavaScript (no framework, no build step)
Embedded	MicroPython on ESP32 CYD, LVGL 9, ILI9341 display, XPT2046 touch

Architecture

Three independent "lines" connected only through HTTP + JSON:

       Web Input  /  CYD touch
              │
              ▼
   POST /analyze_text  ──►  FastAPI (backend/main.py)
              │
              ▼
   nlp_engine.analyze_text()
   tokenize → clean → keyword extraction
            → context classification
            → phrase extraction → summary
              │
              ▼
   SQLite (scenelingo.db)
              │
              ▼
   GET /dashboard  ──►  Web Dashboard  /  CYD Dashboard

API surface:

POST /analyze_text       analyze a text and persist the result
GET  /contexts           list real-world and story contexts
GET  /context/{id}       one context's seed keywords + phrases
POST /save_item          save a word or phrase to the review list
DELETE /saved_item/{id}  remove a review item
GET  /dashboard          aggregated review state (top context, recent keywords, saved items)

Run locally

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
uvicorn backend.main:app --reload

Backend serves at http://localhost:8000. Then open web/index.html directly in a browser, or flash embedded/scenelingo_dashboard.py onto a CYD board.

Project layout

backend/        FastAPI routes, rule-based NLP engine, SQLite storage
web/            single-page browser input + dashboard
embedded/       MicroPython LVGL UI for the ESP32 CYD
docs/           product walkthrough, integration log, per-line task records
img/            screenshots and reference photos

More docs

• docs/Scene_Words_产品介绍.md — full product walkthrough (Chinese)
• docs/INTEGRATION_RECORD.md — three-line integration log
• docs/LINE_A / docs/LINE_B / docs/LINE_C — per-line task records

Team

Built at the Aberystwyth University hackathon by:

• @Geist-L — Line A (FastAPI backend, rule-based NLP engine, SQLite layer)
• @ava-wei666 — Line C (CYD embedded UI in MicroPython + LVGL), Web–CYD integration
• Ruotong Huang — Line B (web input page, web dashboard)