Seismo: a DIY seismic station

A single-channel, sensitivity-first seismometer in a garage in Oakmont, Santa Rosa, California, on the Rodgers Creek / Maacama fault system. Recording continuously since 2026-07-20 as station SS.OAKM1.00.EHZ. Live at https://seismo.mcguinness.ai.

It started with a chandelier at home that I noticed would swing when the ground shook, and wondering if it would make a good improvised seismometer. It didn't. But I progressed to a Raspberry Pi, a 24-bit ADC and a 4.5 Hz geophone that could. Here's a first prototype; the final version is essentially the same, just done a lot neater:

Does it work? Does it work well? To answer that, its findings are checked against a USGS National Strong Motion Project accelerometer 1.64 km away, and after one empirical sensitivity factor the two agree well enough to validate the system.

The system in three boxes

1. The station (Raspberry Pi 2B, garage): geophone, ADS1256 digitizer, miniSEED day-files at 100 sps, timed from a GPS stratum-1 clock on the LAN. It does nothing but acquire; a busy Pi is a noisy Pi.
2. The data engine (Raspberry Pi 5, on the house LAN): the owned archive, re-detection, a gradient-boosting classifier that scores every trigger, a JSON API.
3. The public website (cloud VPS): the dashboard, fed outbound-only from the house.

Read on

• The story: why, how, and what a hobbyist can and cannot claim.
• Architecture: the three hosts, why they are three, and how this differs from a Raspberry Shake.
• The trigger classifier: how the station decides an alarm was an earthquake and not a truck, from STA/LTA to gradient boosting, at undergraduate level.
• Reproducing the results: the comparison against NP.1835, the detection-range fit, and retraining the trigger classifier.
• Catches: every confirmed earthquake beside the professional station's record, on the same axes.

Navigating the repo

STATUS.md is the running log, newest first, with the current-system summary at the top; STATUS-ARCHIVE.md holds everything before 2026-08-20. BACKLOG.md is deferred work. specification.md is the original design and the alternatives rejected. CLAUDE.md is the working brief for the AI side of the collaboration and doubles as the best map of where the code runs.

MIT licensed. Not a professional instrument; treat every number here as a hobbyist's, checked against a professional station where it could be.