Skip to content

Releases: eulogep/Engineer-Learning-OS

v0.3.0 — On-device speech recognition

Choose a tag to compare

@eulogep eulogep released this 22 Oct 17:45
faf48e8

Live demo: https://daily-english-mission-ten.vercel.app

Highlights

🎙️ Speak, and the browser listens — no cloud (#9)

  • On-device transcription: Whisper runs inside your browser (WebAssembly, Web Worker). Record your spoken English and it is transcribed in seconds, ready to review before correction.
  • Pronunciation feedback: read the corrected text aloud; each word is checked against what was heard, with a score and the unrecognised words highlighted.
  • Private by design: off by default and enabled with one click. Your voice never leaves the device; the only network access is a one-time ~77 MB download of the open-source model (Xenova/whisper-base.en, Apache-2.0) from Hugging Face, cached afterwards (works offline). The ONNX runtime is served by the app itself, no CDN code.

💾 Keep your progress (#10)

  • Export / import your Daily English progress from the Données page. Imports are validated strictly and merged, never overwritten.

🛠️ Quality

  • CI on every pull request (#8): lint, typecheck and the full test suite, including real-Chrome integration tests.
  • Local correction now also catches "I am agreed", a frequent transcription of "I am agree" (#11).
  • Hardened Vercel allowlist (#6) and a more robust Chrome test harness (#12).

Quality: 565 tests · typecheck · lint · production build · verified end to end on production.

🤖 Generated with Claude Code

v0.2.0 — Local-first Daily English & spaced review

Choose a tag to compare

@eulogep eulogep released this 10 Oct 13:56
ddf85c4

First public release of Engineer Learning OS: an evidence-grounded, local-first learning OS for engineers.

Live demo: https://daily-english-mission-ten.vercel.app — no sign-up, data stays in your browser.

Highlights

Daily English Mission

  • Works anywhere: falls back to a browser-only mode when no database is available (hosted demo), with a "Mode local" badge (#4).
  • No external AI required: on-device recording for replay only, typed production, deterministic correction of frequent French-speaker errors with explanations in French (#4).
  • Spaced review: Leitner boxes (1 → 3 → 7 → 14 → 30 → 60 days), 3 daily slots reserved for new words, packs interleaved; no schema migration (#4).
  • 150-word vocabulary in 5 packs — essentials, networking, cybersecurity, meetings, career — with an example sentence for every word, shared by the SQLite setup and browser mode (#4, #5).
  • SQLite mode verified on real databases; fixes a mission lasting two days, unstable word order on reload and new words shown as reviews (#5).

Safety & deployment

  • NotebookLM automation bridge disabled by default; requires an explicit server-side opt-in (#2).
  • Hardened Vercel deployment: allowlist .vercelignore, so secrets, local databases and private knowledge are never uploaded (#3).
  • Reproducible local workflow, self-contained npm test (#1).

Upgrading a local install

npm ci
npm run db:setup:local   # adds the 120 new words; keeps your history

Quality: 552 tests pass · typecheck · lint · production build.

🤖 Generated with Claude Code