AI / Automation
Gemma 4 RSS Intelligence Agent
RRRTX Systems — Open Source
At a glance
25+ feeds every 6 hours · $0 monthly AI cost · $7/month total infra · 128K-context local model
The challenge
Staying current in software means watching two dozen or more sources — engineering blogs, release notes, industry news — and most of what they publish is noise: reposts, listicles, tutorial filler. Reading it all costs hours a week; skimming it means missing real signal. The obvious fix — piping every headline through a cloud LLM API — has two problems: at 25+ feeds checked continuously, classification runs $15–20 per month in API spend, and every headline leaves your infrastructure on someone else's.
The system
We built a self-hosted intelligence agent instead: a Python service that polls 25+ developer RSS feeds every six hours, feeds each batch to Gemma 4 E4B running locally through Ollama, and asks the model to do what a good research assistant does — separate the signal from the noise and summarize what actually matters. Gemma 4's 128K context window lets entire batches be classified in a handful of calls rather than one call per headline, which keeps a full batch down to seconds on modest hardware. Clean, summarized digests land in Slack; everything else is discarded. The whole system runs on a $7/month VPS with 3GB of RAM. Feeds, keywords and Slack routing are configured in a single config.yaml, deployment is a one-line install script, and the project is open-sourced under Apache-2.0.
The result
Monthly AI cost: $0 — the model runs on hardware we already pay for. Total infrastructure spend is $7/month, versus $15–20/month for the equivalent OpenAI-powered pipeline, with no per-request billing and no data leaving the VPS. The agent classifies a full batch in roughly 4–8 seconds, posts clean Slack digests on a six-hour cadence, and has replaced manual feed reading entirely. The code is public — monitor.py, config.yaml and install.sh — Apache-2.0 licensed and reusable by anyone with a small VPS and the same noise problem.
Have a system like this to improve?
Start with the problem, constraints, and outcome. We will map the safest next step.
Discuss Your Project