AralOffline: Tagalog tutor that runs fully on a ₱5,000 Android phone

Problem Statement

In many GIDA and island schools, internet is absent or too expensive for daily use, yet most edtech assumes always-online cloud APIs. Teachers handle 40-60 pupils with no aide, and pupils who fall behind in reading never get 1:1 practice. Parents share one low-end Android phone with no load. Any solution that needs accounts, OTPs, or cloud inference fails on day one.

Proposed Solution

AralOffline ships as a single APK sideloaded via USB/SD card. It covers DepEd MELC reading and math for Grades 4-6: phonics practice, Tagalog reading passages with oral scoring, and math drills with step hints. All inference runs on-device (quantized LM + Vosk STT). No account, no personal data collection. Progress stays in local SQLite; teachers export a class summary via QR code scanned from their own phone — no server round-trip.

Target Market

• Primary users: Pupils Grades 4-6 in off-grid/low-connectivity schools, ALS learners.
• Deployers: Teachers, PTA officers, parish/mission schools, NGOs doing last-mile education.
• Scale path: Division offices can preload tablets; youth volunteers can sideload during brigada.

How AI Is Used

• On-device small LM: Generates leveled Tagalog passages and hints. Quantized so it runs on 3GB RAM phones.
• Offline speech scoring: Vosk-based Filipino acoustic model scores oral reading (fluency, missed words) without uploading voice.
• Adaptive drilling: Simple mastery model picks next items; runs locally, no tracking server.
• Why AI: Rule-based drills can’t generate fresh passages or score speech; cloud LLMs can’t work offline.

Independent Deployment Angle

• No telco, cloud, venue, or AR hardware needed. Works on any phone, any SIM or no SIM.
• Distribution is sneakernet: APK on USB, Bluetooth share, school laptop.
• Maintenance by teachers, not a NOC. Open question bank any division can fork.

Demo Scenario (3 minutes for judges)

1. Phone in airplane mode. Open AralOffline, pick “Pagbasa — Baitang 4.”
2. Pupil reads a Tagalog passage aloud; offline STT highlights missed words and gives a fluency score — all on-device.
3. Show a math drill where the tutor gives a hint in Tagalog, not just the answer.
4. Teacher scans the pupil’s progress QR with a second phone; class summary appears with no internet.
5. Close: “Airplane mode on. Learning on.”

30-Hour Build Plan

Suggested stack: Flutter + ONNX Runtime Mobile, quantized Tagalog TinyLM, Vosk offline STT, SQLite, open DepEd MELC question bank. Public data / APIs: DepEd MELC guides (public), open Tagalog corpora, Vosk models.

Risks & Privacy Notes

• Child safety: No accounts, no chat with strangers, no data leaves the device. Complies with Data Privacy Act data-minimization; nothing to retain or breach.
• Competition ToS: Demo uses synthetic pupil profiles only. No real minor’s name, photo, or voice uploaded to any third-party AI tool.
• IP: Question bank is team-written or public-domain; Vosk/ONNX licenses listed in credits.
• Accuracy: Speech scoring shows confidence and lets teacher override; never auto-grades report cards.

AI Use Disclosure (draft)

AralOffline uses a quantized on-device language model for passages/hints and Vosk offline speech recognition for oral scoring. AI coding assistants helped scaffold Flutter screens. All learning content was reviewed by the team against MELCs; demo scores were checked by a human.

Rubric Scoring