A home network enthusiast fed 650K+ DNS logs from Pi-hole and AdGuard Home into Gemma 4 (running locally via Ollama) to see what patterns a local LLM could surface. Because the raw CSV was far too large for any model's context window (~12M tokens vs 128K limit), the author pre-aggregated the data client-side into a structured JSON summary using PapaParse, then forwarded it to Gemma via a lightweight Python proxy and a React/FastAPI dashboard. Without being given any device names or IP labels, Gemma correctly grouped Amazon domains and attributed them to a Fire TV stick, identified a 4–7 AM traffic spike tied to Nextcloud syncs and background jobs, and flagged idle devices continuously phoning home — none of which the standard Pi-hole or AdGuard dashboards had surfaced. The key insight: the intelligence came as much from thoughtful data preparation as from the model itself.

8m read timeFrom xda-developers.com
Post cover image
Table of contents
The data was messier than I expected677,000 rows is too many for a promptWhat Gemma found that I missed
77 Impressions