Tilmidh Agent (تلميذ)

Inspiration

Every academic year, millions of Moroccan parents and primary school students (1AP–6AP) face an invisible mental load: tracking scattered continuous assessment schedules, hunting for regional exam PDFs across social media groups, and staying aligned with weekly textbook breakdowns (Sommaire).

Existing educational platforms demand account creation, phone numbers, and constant manual searching. We wanted to build a Good Neighbor Agent that runs silently in the background with zero friction—an offline-first assistant that needs zero authentication and surfaces actionable context cards only when real attention or review is required.

What it does

Tilmidh Agent (تلميذ) operates as an autonomous background companion and interactive tutor via a zero-auth Progressive Web App (PWA):

• Autonomous Holiday & Exam Alerts: Evaluates current dates against official Ministry of National Education (MEN) academic calendars to deliver proactive Pre-Holiday Notice Cards (e.g., 7-day warnings before Mid-Term breaks with 15-minute quick-review topics).
• Weekly Sommaire Sync: Ingests official textbook tables of contents across Arabic, French, Math, and Science, organizing weekly subjects automatically into visual cards.
• Exam PDF & Practice Fetcher: Retrieves regional practice exams (e.g., 6AP Examen Normalisé Local/Régional) directly into the browser for 100% offline study.
• Instant AI Primary Tutor: When students or parents open a review card or exam sheet, an integrated AI tutor steps in to explain difficult problems step-by-step in Arabic, French, or Darija without requiring logins or passwords.

How we built it

We prototyped system prompts, tool schemas, and multi-subject Sommaire JSON structures in Google AI Studio to iterate rapidly and ensure structured outputs before moving to cloud production execution.

For production, we integrated the Strands Agents SDK with Amazon Bedrock (Claude 3.5 Sonnet) as our primary reasoning engine. The PWA dispatches anonymous, non-PII payloads (grade, weekNumber, currentDate) to the serverless Strands Agent environment. The agent executes deterministic tools (holidayAlertTool, curriculumScheduleTool, examPdfTrainingTool) and returns structured UI cards and tutoring responses back to the client.

Challenges we ran into

Balancing complete student privacy with context-aware tutoring was our biggest challenge. Because we strictly refused to store user accounts or personal identifiable information (PII), we had to architect the agent to be completely stateless—relying on local browser storage (IndexedDB) to persist downloaded exam PDFs while keeping API payloads strictly anonymized.

Accomplishments we're happy of

• True Zero-Auth Architecture: Parents and children can install and use the PWA instantly without creating passwords or sharing personal data.
• Deterministic & Hallucination-Free: By binding verified curriculum datasets to Strands tools, textbook schedules and exam practice links remain 100% accurate.
• Offline Resilience: Pre-cached exam PDFs and lesson summaries remain fully accessible even in rural areas with low connectivity.

What we learned

Designing agentic AI for families requires shifting from open-ended chat interfaces to context-first UI cards. Minimizing cognitive friction for non-technical parents matters far more than presenting an unconstrained chatbot.

What's next for Tilmidh Agent

We plan to expand tool coverage to secondary education levels (Collège / Lycée) and add voice-first lesson summaries to assist illiterate parents in guiding their children's weekly schoolwork.

Built With

• amazon-bedrock
• claude-3-5-sonnet
• google-ai-studio
• indexeddb
• pwa
• strands-agents-sdk
• typescript