AI / hardwarephase 0 · the plan is built, the farm is not

One Acre, Zero Dependency.

Off-grid, AI-driven farming on two plots 2,256 km apart: Prora through summer, Castelo Branco through winter. One question: how much of a small farm's daily decisions can genuinely off-grid AI handle, from sensor to scheduler?

Open the repo Project brief (PDF)

System architecture: a Pico 2 W field node sending readings over BLE or LoRa to a Python stack on a laptop that ingests, stores, models and schedules — From sensor to scheduler. Storage is the seam.

Two plots, one season

Prora runs through the summer, Castelo Branco through the winter. Between them there is never a month with nothing growing.

Nothing leaves the acre

Sun to sensor to laptop to schedule. The model runs locally, which is the only way the off-grid claim survives contact with a bill.

The model proposes, Python disposes

Every suggestion passes a deterministic rule set before it becomes a task. What fails the rules is dropped, and the drop is logged.

~/what it does

Sensor, history, calendar, model.

A solar-powered field module (Birdbox) collects soil and local weather data. The Open-Meteo historical archive adds twenty years of regional weather. Eurostat adds regional yield figures, coarse ones: annual, per hectare, per region, not per bed. The Thun calendar supplies a moon-based planting calendar, which this project computes correctly and tests instead of assuming.

The calculations run in Python. A local model lays the results side by side and proposes which crops give the best yield per planting, when to sow, and which rotations work on 4,000 m².

The aim: something farmers, hobby growers and families can hold on to, to grow organically at any scale, on modern technique and age-old concepts that protect the earth.

~/where

Two plots, 2,256 km apart.

14.6 degrees of latitude apart. You follow the growing season instead of fighting it.

Prora summer

where: Rügen, German Baltic coast
latitude: 54.39 °N
reference: DWD Arkona, station 00183, 33 km north
winter solstice: 7.0 h of daylight, sun peaks at 12.2 °

Castelo Branco winter

where: Central Portugal, inland
latitude: 39.82 °N
reference: IPMA long series, plus ERA5-Land
winter solstice: 9.2 h of daylight, sun peaks at 26.7 °

Prora leans on DWD Arkona, with measured radiation from 1947. Castelo Branco has no equivalent open history and leans on ERA5-Land. That asymmetry is not papered over.

~/commitments

Three things it does not compromise on.

Off-grid means off-grid.

No cloud LLM in the daily loop.

Every number carries its error bars.

A €30 sensor claiming ±0.3 pH does not get to report pH to one decimal.

The calendar is a hypothesis.

The biodynamic calendar is under test, not a scheduling rule.

~/the field node

A Pico, not a Pi.

Tap a picture to see it full size.

Winter solstice sun at the two plots: Prora 7.0 hours of daylight with the sun peaking at 12.2 degrees, Castelo Branco 9.2 hours at 26.7 degrees — Prora gets 7.0 hours of daylight at the solstice, sun at 12.2 degrees. That is what the node is sized for.

Daily energy comparison: Pi always-on 14.4 Wh, Pi duty-cycled 2.4 Wh, Pico W duty-cycled 0.16 Wh — A duty-cycled Pico runs on roughly sixty times less than an always-on Pi.

Cross-section of the field node: birdbox shell over an IP66 box, vent and glands on the underside, radiation shield on a tail, panel at 65 degrees, soil probes at 10 and 30 cm — Two failure modes kill the node sooner than any sensor: condensation, and charging below zero.

Link choice by distance: BLE under 20 metres, marginal from 20 to 50 metres, LoRa beyond 50 metres — Under 20 metres BLE, above that LoRa. The ingest layer makes that a one-module change.

~/the calendar

The Thun calendar, honestly implemented.

The evidence does not support the trigon effect. The calendar is therefore logged as a covariate and tested, never used as a rule.

Ring chart of the time the Moon spends in each constellation over one sidereal month, coloured by Thun day-type, with Ophiuchus and non-zodiacal strays marked — Computed from the ephemeris in thirty-minute steps. Sidereal, not tropical.

A tropical sign band above a shifted, unequal constellation band; the same Moon position reads as Gemini tropically and Taurus in the real sky — Tropical signs sit 24 to 25 degrees off the real constellations. Get this wrong and every day type is wrong.

Real IAU boundaries are unequal, so 2027 comes out at 107 root days against 67 flower days.

Two sine waves at 27.32 and 29.53 day periods drifting apart, from half a day at the first crest to 2.8 days at the second — Ascending and descending is a 27.3-day cycle, entirely separate from waxing and waning at 29.53.

~/models

The LLM proposes, Python disposes.

Every number is computed in tested Python. The model selects, prioritises and explains, and returns JSON validated against a schema. What fails the rules is dropped, and the drop is logged. What it is aimed at: yield per square metre.

~/roadmap

Ten phases.

~/see the real thing

Read the plan. Read the code.

CODE

AI-Homestead

The repo. Build plan, ADRs, data sources, and the figures on this page.

PDF

Project overview

Short brief. The two plots, the three commitments, and what each data source can and cannot claim.

STACK

Pico 2 WMicroPythonPython 3.11SQLiteskyfieldLSTMOllamaBLE / LoRaDWD / IPMAERA5-Land

MIT for the code. Weather data from DWD (CC BY 4.0), IPMA, and Open-Meteo (CC-BY-4.0). Ephemerides from JPL via skyfield. No copyrighted calendar data is redistributed.

~/next

Same rigour,
your problem.

--:-- in Kaunas · office hours · Mon to Fri, 10:00 to 18:00

Start a project All builds