Tutor sa Bulsa: An AI tutor that works over plain SMS or fully offline, so no learner is left behind by bad signal

Problem Statement

Philippine learners scored among the lowest of participating countries in PISA 2022, and the World Bank has estimated very high learning poverty: most 10-year-olds can’t read and understand a simple text. Many Geographically Isolated and Disadvantaged Areas (GIDAs), island barangays, and typhoon-hit communities have weak or no mobile data, so app-based and video tutoring don’t reach them. Basic feature phones and SMS, however, are nearly everywhere. Typhoon class suspensions and modular learning leave children studying alone with parents who may not be able to help. Private tutoring costs money most families don’t have.

Proposed Solution

Tutor sa Bulsa has two delivery modes that share one curriculum brain. SMS mode: a learner texts a keyword (for example, “MATH4”) to a short code and chats with an AI tutor that asks Socratic questions in short messages of up to 160 characters, gives hints rather than answers, and serves quick quizzes. No data plan or smartphone is needed. Offline mode: an Android app bundles a small quantized language model (around 1–3B parameters) plus a compressed bank of DepEd competency content, so tutoring works with zero signal on budget phones and syncs progress when the phone reconnects. Both modes follow the DepEd most essential learning competencies, switch between Filipino, English, and local languages, and adapt difficulty to the learner. Teachers receive a weekly mastery summary by SMS or on a dashboard.

Target Market

• Primary users: Grade 4–10 learners in GIDAs, island and upland schools, and ALS learners, potentially millions of students in under-connected areas.
• Secondary users: teachers, parents, and barangay learning hubs.
• B2G buyers: DepEd (learning recovery and ALS programs), LGU Special Education Funds, DICT. CSR and partners: PLDT-Smart Foundation, UNICEF, and education NGOs.

How AI Is Used

• Small on-device LLM (quantized, e.g., Gemma 2B or Qwen 1.5B class): delivers tutoring dialogue with no connectivity. Pre-written content can’t respond to a child’s specific confusion.
• Cloud LLM in SMS mode: gives richer Socratic tutoring compressed into short texts. A message-length-aware prompt keeps each reply under 160 characters.
• Retrieval over the DepEd competency bank: keeps explanations aligned with the curriculum and reduces hallucination.
• Knowledge tracing (lightweight Bayesian or Elo model): estimates mastery per competency and picks the next question. This is what makes it a tutor rather than a chatbot.
• Language detection and translation: answers in whatever language the learner uses, including code-switched Taglish.

PLDT / Smart / ePLDT / Rokid Angle

• Smart SMS short code and A2P messaging: a core PLDT revenue rail, which could be zero-rated or toll-free for learners as a CSR or B2G offering.
• ePLDT cloud: hosts the SMS tutoring LLM, learner profiles, and teacher dashboards in Philippine data centers.
• Smart network expansion: as GIDA sites get coverage, learners upgrade smoothly from SMS to app mode.
• PLDT-Smart Foundation tie-in for school pilots. Rokid (stretch): AR worked examples for teachers in learning hubs.

Demo Scenario (3 minutes for judges)

1. Show a photo of a Grade 5 learner, “Junjun,” on an island in Romblon. Typhoon signal no. 2 is up and classes are suspended.
2. On a basic keypad phone on stage, he texts “SCI5” to the short code. The AI replies in Filipino: “Kumusta! Ano ang alam mo tungkol sa tatlong anyo ng matter?” (What do you know about the three states of matter?)
3. He answers wrongly, and the tutor gives a hint using yelo (ice) and singaw (steam) instead of the answer. He gets it on the second try.
4. Switch the Android phone to airplane mode. The offline app explains fractions in Romblomanon-flavored Filipino using a local model.
5. Reconnect: the progress syncs. The teacher dashboard shows class mastery and flags three learners stuck on fractions.
6. Close: “Kahit walang signal, may tutor.” (Even without signal, there’s a tutor.)

30-Hour Build Plan

Suggested stack: Kotlin or Flutter Android app with llama.cpp or Google MediaPipe LLM Inference, Gemma 2B or Qwen2.5-1.5B (int4), SQLite with a small embedding index, FastAPI with a Gemini or GPT-class LLM for SMS mode, a Semaphore or Twilio SMS gateway for the prototype (Smart A2P in production). Public data / APIs: DepEd MELCs and MATATAG curriculum guides, DepEd Commons self-learning modules (check licenses), PSA OpenSTAT, DepEd school data on data.gov.ph (to map GIDA schools), the NTC or OpenCelliD coverage dataset, and Hugging Face open models and Tagalog datasets.

Risks & Privacy Notes

• Small-model quality and hallucination: limit scope to curriculum retrieval, require hint-first answers, use human-verified item banks, and escalate to “ask your teacher” when the model isn’t confident.
• Phone performance and battery: use int4 quantization, short contexts, and target phones with 4 GB of RAM or more for offline mode. SMS mode covers the rest.
• SMS costs: need a toll-free or CSR-funded short code, and cap the number of messages per day.
• Data Privacy Act 2012: learners are minors. Get parental consent, store only a mobile number hash and mastery scores, keep data on-device by default, host in PH, and give teachers access only to their own classes.
• Clause 24 compliance: the demo uses test numbers and synthetic learner profiles. All AI-generated content and models are disclosed.

AI Use Disclosure (draft)

We use an open small LLM (e.g., Gemma 2B or Qwen 1.5B) on-device and a Gemini or GPT-class LLM in the cloud for tutoring dialogue, plus embeddings for curriculum retrieval. An LLM helped draft quiz items, which teammates checked against DepEd competencies. Coding assistants helped with boilerplate, and an LLM helped edit this proposal. No real student data is used, and humans review all teaching content.

Rubric Scoring