04 / Farm operations
Ruralist
Fieldhaus
The problem
A small farm runs a dozen businesses at once, each with its own calendar, records and money. The tools for each are separate, and most of them are built for much bigger operations.
An assistant that can change farm records on its own is a risk. The farmer has to stay the one who decides.
The data is wide as well as deep. One account owns several sites, each enterprise keeps its own calendar and records, and a site's profile is assembled from public sources for parcel, terrain, sun, soil and climate, each with its own date and terms of use. Money has to be valued the same way on every screen, or the morning view and the ledger will disagree.
It also had to run for someone with no keys and no setup, so that a farmer, or anyone reviewing the work, could open it and see the whole product with a crew that still reads real records.
How it was built
The first design decision was about trust. The farmer stays the one who decides, so the crew proposes and never edits. The look followed from the setting: a field station's console at night, with instruments that draw in as their readings arrive.
- 01
Today first
The home screen answers the four morning questions on one page, with a live sun arc in the top bar. Everything else is a tap away, and a command palette, Ask the crew and a quick-log sheet are available everywhere.
Today, on the fictional demo farm. - 02
A site from public data
Add a site by address or by drawing its boundary, and a pipeline builds a layered profile from public data: parcel, terrain, sun and shade through the season, structures and canopy, soil and climate. Each layer carries its source and date and can be corrected.
A site, before and after its public-data profile. - 03
Land and spaces
Beds, fields and indoor spaces are drawn over the site, and indoor rooms carry their own readings.
Land and spaces, drawn over the site. - 04
A pack per enterprise
Each enterprise, from vegetables to bees to the woodlot, is a pack with its own dashboard and tools, built on one engine so a new pack adds screens without new plumbing.
Every enterprise, with its own dashboard. - 05
A crew that proposes
A foreman routes each request to specialists such as a rotation planner, a mycologist, a stockkeeper and a bookkeeper. They run in parallel, read the records through tools and answer with the numbers they used. They never change data directly: changes arrive as proposals.
The crew, and proposals waiting for a yes. - 06
Honest value
A savings ledger counts spoiled harvests as zero and sold ones at the price received, shows labor separately, and says so when growing cost more than buying.
The savings ledger, with honest math.
The solution
A Next.js 16 app in a Turborepo, with the domain logic, contracts, packs and design tokens in packages that a React Native app can reuse. It installs as a PWA.
Out of the box it runs with an embedded database, a demo sign-in and a mock model, so the whole product works with no keys. With an API key the crew runs on live models, capped at $1 a run and $25 a month by default.
Decisions along the way
Proposals, never edits
The crew reads records through tools and answers with the numbers it used. It cannot change data. Every change arrives as a proposal the farmer accepts or rejects.
Honest math
A spoiled harvest counts as zero, a sold one at the price received, and labor is shown apart. When growing cost more than buying, the ledger says so.
Charts are laid out, not scaled
A scaled drawing shrinks its lettering to 7 pixels on a phone. Charts that sit on a grid are built as HTML with fixed-size type, and one hover readout serves every mark.
Respect the map's terms
Public-domain imagery is toned for night so the glowing layers read. Google's tiles are left untouched, because their terms do not allow altering them.
What I delivered
- Today, Sites, Land, Plan, Enterprises, Soil, Harvest, Inventory, Money, Sales, Journal, Value and Crew screens
- A site pipeline over public data
- Enterprise packs on one engine
- A crew of agents that proposes changes
- A savings ledger and evidence library
- About 600 unit and API tests, plus browser tests of the crew
Value added
For the farmer
One morning view of the whole place, and help that explains itself and waits for a yes.
For trust
Money is valued the same way on every screen, and the savings ledger says so when growing cost more than buying.
By the numbers
US farms by size, 2024
Share of all farms beside share of the value of production. Small family farms have gross cash farm income under $350,000.
- Share of farms
- Share of production value
View as table
| Item | Share of farms | Share of production value |
|---|---|---|
| Small family farms | 86% | 17% |
| Midsize family farms | 6% | 18% |
| Large-scale family farms | 5% | 50% |
| Nonfamily farms | 3% | 14% |
Source: USDA Economic Research Service, Charts of Note, May 28, 2026 (opens in a new tab)
The result
A public demo of the fictional farm runs at ruralist.vercel.app, with no accounts and no keys; the crew answers from the built-in mock model.
Results
- ~600unit and API tests
- 4questions answered every morning
- $1default cap on a live crew run
What I took from it
The design call that mattered most was making the crew propose instead of act. It kept the farmer in charge and made the whole product easier to trust, and it made the mock model useful too: a proposal can be judged the same way whoever wrote it. It also made the crew easy to check. The mock files real proposals through the same seam as a live model, so the whole flow can be tested end to end without a key.
Background
The audience is larger than it sounds. The USDA's Economic Research Service counts small family farms, those with gross cash farm income under $350,000, as 86 percent of all U.S. farms in 2024. They operated 40 percent of farmland and produced 17 percent of the value of production. Tools for farm records are mostly built for the other end of that scale.