Production AI & Backend Automation • Self-Hosted 24/7 Daemon

Classroom Sentinel — 24/7 Autonomous Python Daemon with Schema-Constrained Gemini Deadlines

A self-hosted worker that turns classroom announcements into verified, idempotent deadline reminders.

PythonasyncioGeminiPydanticTelegram

Role: Solo backend and AI automation engineer
Production status: Self-hosted 24/7 production Linux daemon
Core stack: Python asyncio · aiosqlite · Google Classroom API · Gemini · Pydantic · Telegram Bot API · SQLite
Business outcome: Built an unattended worker that monitors classroom feeds across two Google identities, extracts deadlines, and sends duplicate-safe reminders.

The real-world problem

What the system had to solve

University instructors post assignment deadlines inside casual, unstructured conversational announcements in both Arabic and English. Students frequently miss submission deadlines because basic prompt wrappers hallucinate dates and crash backend parsers by outputting invalid JSON formats.

What I built

Architecture shaped by the constraints

I built a self-hosted 24/7 background worker using Python asyncio and aiosqlite that continuously polls the Google Classroom API. Unstructured announcements pass to Google Gemini models bounded by strict Pydantic JSON schemas, and validated deadlines trigger stateful, idempotent 24-hour and 2-hour Telegram reminders.

Production proof

Delivered system

Running continuously in production on a self-hosted Linux server, tracking live course feeds and delivering zero-duplicate deadline alerts without maintenance interventions.

System flow

Architecture at a glance

Google Classroom

Course announcements are polled asynchronously.

Gemini + Pydantic

Unstructured posts become typed deadline records.

SQLite + Telegram

Stateful, duplicate-safe reminders are scheduled and dispatched.

Engineering decisions

Why the system works this way

Deterministic fields before AI extraction

Constraint: Sending every assignment through an LLM adds cost and can hallucinate dates already present in structured coursework fields.
What I engineered: Read dueDate and dueTime directly for coursework; only free-text announcements use Gemini constrained by Pydantic fields and urgency tags.
Production outcome: Official assignment dates stay deterministic while buried announcement deadlines can still be extracted.

SQLite state for idempotent reminders

Constraint: Repeated polling encounters the same posts and can send duplicate 24-hour and 2-hour alerts.
What I engineered: Persisted announcement IDs and alert flags in SQLite through aiosqlite, and checked them before scheduling or dispatch.
Production outcome: Suppressed repeat reminders across unattended polling cycles.

Headless multi-account token refresh

Constraint: Interactive OAuth libraries may open a browser when tokens expire, freezing a daemon without a desktop session.
What I engineered: Separated background refresh from interactive authorization and silently refreshed stored credentials for personal and university accounts.
Production outcome: Kept the worker running headlessly across both identities, with an operator alert when manual reauthorization is needed.

Discuss a similar system

Product walkthrough

From classroom feed to a reminder

Step 1 of 4

Poll both accounts

An asyncio loop polls Google Classroom course feeds for the personal and university identities and checks stored announcement IDs.