Gemma Monitor — case study

claims → status → evidence — every row opens something you can check yourself
claim	status	evidence
Runs production news-intel for $7.40/month versus the $22/month API subscription it replaced	MEASURED	github.com/ahmadrrrtx/Gemma-4-RSS-Intelligence-Monitor ↗
Structured INCLUDE/EXCLUDE screening prompts so the model triages instead of rambles	SHIPPED	monitor.py ↗
Full write-up with the benchmark tables on dev.to	PUBLISHED	dev.to/ahmad_rrrtx/i-ran-gemma-4-on-a-7month-server-and-built-an-ai-powered-news-monitor-that-costs-0-to-operate-5gn0 ↗

01The problem was the invoice

I was paying ~$22/month for an OpenAI subscription mostly to summarize RSS feeds. That's not an AI strategy, that's a tax. The question that made it a project: what is the cheapest thing that can do exactly this job, reliably, unattended? (Small system; short study. Scope discipline is the point.)

02The system

One well-organized monitor.py: fetch feeds → build a triage prompt → Gemma 4 E4B via Ollama → post digest to Slack. Scheduled by cron, installed by one script, zero API keys. The interesting part is that the model only does the one step it's genuinely good at — ranking signal against a rubric — while fetching, deduplication, and delivery are boring Python. Defensive HTTP handling and logging are in because a headless automation nobody watches is a lie that pings once a week.

Gemma-4-RSS-Intelligence-Monitor/monitor.py the actual screening rubric — triage as a decision list, not a temperature roll view on GitHub ↗

INCLUDE:
- Stable releases of major projects (v1.0+, security patches, major versions)
- Breaking changes or deprecations in popular frameworks
- Critical security vulnerabilities (CVE, patches)
- Significant new capabilities or architectural changes
- Major ecosystem announcements

EXCLUDE:
- Tutorials, how-to guides, "getting started" content
- Promotional blog posts, sponsored content
- Minor patch releases (unless security-related)
- Opinion pieces without new technical information
- Duplicate news from the same event

INCLUDE: - Stable releases of major projects (v1.0+, security patches, major versions) - Breaking changes or deprecations in popular frameworks - Critical security vulnerabilities (CVE, patches) - Significant new capabilities or architectural changes - Major ecosystem announcements EXCLUDE: - Tutorials, how-to guides, "getting started" content - Promotional blog posts, sponsored content - Minor patch releases (unless security-related) - Opinion pieces without new technical information - Duplicate news from the same event

03The numbers (method: the README's benchmark tables)

line	$22 baseline	this
monthly cost	$22.00 API subscription	$7.40 (all of it server)
API dependency	yes, key required	none — local inference
model trade-offs	—	2B / 9B / 27B measured, tabled, chosen

66% reduction, reproducible by running the benchmarks on the same feeds for a week. That sentence, with that method line, is why this small repo out-punches big ones.

04Why this stays on the site next to a runtime and a GPT

Because it's the artifact I'd want to be interviewed about: a genuinely small problem, fully owned, defended by numbers. The write-up on dev.to carries the same tables — writing it forced me to keep the claims and the code consistent.

Limitations, stated plainly

"Running in production for weeks+" is author-reported (mine) — the box is mine and private; the code and benchmark method are public.
No test suite here; a cron script of this size doesn't earn one, and I won't pretend it does.
Slack delivery is hard-wired for my use; a general tool would need a sink abstraction.
A second contributor later edited the README — history's public, credit's in it.