TECH

Building PiSchool: An Automated School Communication Pipeline with Local LLMs

26 June 20268 min read

AIAutomationPythonRaspberry PiLLMGmail APIDevOpsSystem Design

The Problem

My children attend two different schools, each using its own communication channels — one via BromCom, another via Studybugs, plus direct emails from both. Every week brings a firehose of emails: early closures, PE kit reminders, Finham Fest wristbands, heatwave guidance, SEND parents' evenings.

Critical information gets buried. An early-closure announcement at 8:15 AM might sit unread until lunchtime. A reminder to send a water bottle gets lost between newsletters. I found myself manually scanning Gmail every morning, mentally piecing together a timeline of what's happening and what needs packing.

Existing solutions (school apps, parent portals) are school-specific. There's no cross-school aggregator — and certainly none that works with both state schools in Coventry simultaneously.

The Constraint

I'm not paying for an API. No OpenAI credits, no Anthropic tokens, no cloud LLM bills. The extraction has to run locally, for free, on hardware I already own.

The Solution: PiSchool

An 8-phase pipeline that costs zero per run:

┌─ Windows Dev Machine (Task Scheduler · Mon-Fri 10am) ─┐
│  1. Fetch → Gmail API (readonly scope)                 │
│  2. Extract → Ollama + Qwen2.5-Coder:1.5B (local LLM) │
│  3. Push → SCP to Pi 5 (Nginx serves JSON)            │
│  4. Notify → Gmail API (send scope) → email digest     │
│                                                        │
│  → sanarahman1@gmail.com                               │
│  → farhanauddin208@gmail.com                           │
└──────────────────────────┬─────────────────────────────┘
                           │
┌─ Pi 5 (192.168.1.219) ──▼─────────────────────────────┐
│  Nginx → /admin/data/school-comms.json (6KB JSON)      │
│        → /admin/pages/pischool.html (dark-theme UI)    │
│  256MB MemMax · 1 core CPUQuota · static files only    │
└────────────────────────────────────────────────────────┘

Phase 1: Gmail API Setup

The foundation. A Google Cloud project with the Gmail API enabled, OAuth 2.0 Desktop credentials, and a consent screen configured for external use (with test user access).

The credentials live in Hidden/pischool/ — .gitignore'd, never committed. Authentication is a one-time browser flow that saves a token.json for subsequent runs.

Scopes requested:

gmail.readonly — fetch school emails
gmail.send — send digest emails

Phase 2: Email Fetcher

A Python script (fetch_emails.py) that builds a Gmail API search query from configured sender domains:

SENDER_DOMAINS = [
    "grangefarm.coventry.sch.uk",
    "bromcomcloud.com",
    "info1.studybugs.com",
    "finhampark.co.uk",
]

The query searches for unread emails matching from:(domain) label:"Kids school" within the last 7 days. Each email is parsed (multipart, HTML, plain text → beautifulsoup extraction) and saved as raw JSON to data/raw/{message_id}.json. Deduplication by message ID prevents reprocessing.

First run: 13 school emails captured from Grange Farm and Finham Park.

Phase 3: Local LLM Extraction

The clever bit. Every raw email is sent to Ollama running Qwen2.5-Coder:1.5B — a 986MB quantized model that runs entirely on my Windows dev machine. No data leaves my network.

The system prompt defines a strict JSON schema:

{
  "event_type": "closure | exam | pe_day | award | bring_item | meeting | deadline | other",
  "title": "Event title",
  "date": "YYYY-MM-DD",
  "date_end": "YYYY-MM-DD | null",
  "time": "12:35pm",
  "location": "School",
  "action_required": "What the parent needs to do",
  "urgency": "high | medium | low"
}

The model converts free-form email text into structured events. It handles date ranges ("24th-26th June"), relative dates ("tomorrow", "this Friday"), and correctly classifies urgency (closures = high, newsletters = low). Each extraction takes 8–20 seconds per email.

The first batch of 13 emails yielded 12 structured events — early closures, instrument lesson cancellations, heatwave guidance, PE kit reminders, SEND parents' evening, Finham Fest wristbands. Zero misses.

Phase 4: Pi 5 Deployment

The Pi 5 runs Nginx with strict resource limits (256MB MemoryMax, 1 core CPUQuota). A PowerShell script (push-to-pi.ps1) SCPs the generated JSON over SSH to /var/www/project2jarvis/admin/data/school-comms.json.

The Pi is a passive server — it only serves static files. No Python, no Node, no cron. Just Nginx and a 6KB JSON file.

Phase 5: Automation

A master runner script (run-pischool.ps1) orchestrates all 4 stages with error handling at each step. Windows Task Scheduler fires it Monday to Friday at 10:00 AM.

Stage 1: python fetch_emails.py      # Gmail API → raw JSON
Stage 2: python extract_events.py    # Ollama → structured JSON
Stage 3: .\push-to-pi.ps1            # SCP to Pi 5
Stage 4: python notify_email.py      # Gmail API → email digest

The full pipeline completes in under 3 minutes with 5–10 new emails.

Phase 6: Dashboard

A dark-theme single-page dashboard (pischool.html) served by Nginx. Vanilla JavaScript fetches the JSON and renders:

Summary cards: events this week, urgent items, items to bring, actions needed
Filter pills: closures, exams, bring items, meetings, deadlines
Timeline view: events grouped by date with urgency dots and date badges
Pack checklist: deduplicated bring-items with localStorage persistence

The entire dashboard is 688 lines — no frameworks, no build step, no backend.

Phase 7: Email Notifications

The notify_email.py script builds a plain-text digest of recent events and sends it to both parents via the Gmail API.

Deduplication: A SHA256 hash of (message_id + title + date) is stored. If nothing changed since the last digest, no email is sent. No spam.

Sample digest format:

📋 PiSchool — School Communication Digest
  Friday, 26 June 2026

  🔴  Early Closure (2026-06-24)
       🕐 12:35pm
       📌 Contact school Reception to plan in place

  🎒  Water Bottle and Sunscreen (2026-06-29)
       📌 Send child with full water bottle and sunhat

  🔗  Full dashboard: http://192.168.1.219/admin/pages/pischool.html

First digest sent successfully at 02:18 AM to both recipients.

Phase 8: Documentation

A complete product manual (20260626-PiSchool-Complete-Documentation.html) documents the entire system — PRD, system architecture, technical specs, build manual with all 8 phases, troubleshooting guide, and appendix. Dark-themed, printable, single-file HTML with embedded CSS.

A copy also lives in the digital products folder at digital-products/pischool-manual/.

Key Design Decisions

Local LLM over API: Qwen2.5-Coder:1.5B (986MB quantized) runs on Windows via Ollama. Zero per-run cost, zero data leakage. The trade-off is speed (~15s/email) but the 10am schedule means it runs once daily — speed isn't critical.

Plain-text over HTML emails: Gmail strips JavaScript and most CSS in transit. A plain-text digest with a link to the full dashboard is more reliable and works everywhere.

Two-level deduplication: processed_ids.json prevents re-fetching emails. last_digest.json (SHA256 hash) prevents re-sending digests. Both are persistent across runs.

Pi stays static: The Pi 5 serves only pre-generated JSON and HTML. No runtime processes, no databases, no API endpoints. This keeps it within its 256MB cgroup allocation alongside Minecraft's 3GB allocation.

10am Mon-Fri schedule: School emails arrive in morning batches. Every-30-minute polling was wasteful. Once daily at 10am captures everything and respects weekends.

What It Looks Like In Practice

http://192.168.1.219/admin/pages/pischool.html

A dark-themed dashboard showing a timeline of school events. Four stat cards at the top: events this week (5), urgent (2), items to bring (3), actions needed (4). Filter pills to isolate closures or PE days. A persistent pack checklist with checkboxes that survive page reloads (localStorage).

Below the timeline: "What to Pack" — water bottle, sunhat, PE kit, wristband — each with a checkbox that stays checked until unchecked.

Current Status

All 8 phases are built, deployed, and tested end-to-end. The first automated Task Scheduler run is set for Friday 26 June 2026 at 10:00 AM.

The system has processed 13 school emails, extracted 12 events, served them on the Pi dashboard, and sent the first digest to both parents — costing exactly zero dollars in API fees.

What's Next (V2)

Priority	Feature	Why
High	Parallel LLM extraction	Cut batch time from 95s to ~25s
High	Search/filter by keyword	Find events faster
High	Link to original email	One-click access to source
High	Per-child filtering	Separate Grange Farm from Finham Park events
Medium	HTML email digest	Colour-coded, styled emails
Medium	Health monitoring email	Know if the pipeline fails
Medium	PWA installable dashboard	Add to phone home screen
Future	Multi-family auth	Sell as a product for other parents
Future	Setup wizard	One-command install for new users

Tech stack: Python 3.11+, Ollama + Qwen2.5-Coder:1.5B, Gmail API, Raspberry Pi 5 + Nginx, PowerShell, Windows Task Scheduler, Vanilla JS.

Cost per month: £0.00. The only electricity cost is the Pi 5 (idle, serving a 6KB JSON file) and the Windows dev machine (already running for other work).

Repository: Part of Project2Jarvis — an Obsidian vault and OpenCode workspace on GitHub.