Why We Built the Farmer Hub: AI Decision Support for Small-Scale Farmers

Most of EcoDesign is built for whole-system design — siting a passive solar building, assembling a plant guild, planning a regenerative landscape from a blank site. The Farmer Hub is different. It was built for the farmer who is already on their land, already growing, and standing in a field with a problem that needs an answer today. This article explains why we built it as a separate experience, what the research told us about small-scale farmers, the specific design decisions that followed, and how the six modules address the problems farmers actually have.

The problem: expertise that never reaches the field

A small-scale farmer — someone working anywhere from half a hectare to fifty, often with limited capital, family labour, and local markets — faces an unusually wide range of problems. In a single season they may need a plant pathologist (what is eating this leaf?), an irrigation engineer (how much water, and when?), a soil scientist (why is this patch compacted and yellowing?), an agronomist (what should I rotate to next year?), and a transition advisor (how do I cut chemical inputs without losing the crop and my income?).

Conventionally, that breadth of expertise is expensive, slow to coordinate, and — for most smallholders — simply unavailable. They get a fraction of it: a neighbour's opinion, an input dealer's recommendation (which tends to favour selling inputs), or a web search that doesn't know their soil, their climate, or their season. The expertise exists in the world; it just never reaches the field at the moment the decision has to be made.

That is the gap the Farmer Hub was built to close. But the interesting part is not the ambition — it is what we learned about why previous digital tools have struggled to close it, and what that forced us to design differently.

What the research told us

Before writing a line of the module, we ran a deep research review of smallholder-facing digital agriculture: the personas, the adoption literature, the open-data landscape, and the existing apps. Two findings shaped everything that followed.

The biggest risk is adoption, not model accuracy. A perfectly accurate recommendation is worthless if the farmer never reaches it, doesn't trust it, or can't act on it. The recurring barriers across the literature were consistent: uneven field connectivity, varying digital literacy, gendered and shared-device access patterns, cost sensitivity, low tolerance for setup, and — repeatedly — scepticism toward "remote" or "AI" advice with no visible reasoning. The implication was blunt: optimise the product for access and trust, not for feature count.

Existing tools have two structural gaps. First, many diagnostic apps drift toward treatment-centric, product-led recommendations rather than an Integrated Pest Management (IPM) logic that puts biological and cultural controls first. Second, almost none of them are transition-first: they optimise the farmer's current practice instead of guiding a stepwise, risk-managed move from conventional toward regenerative. Both gaps were also opportunities — places where a guardrail-heavy, transition-aware tool could be genuinely different rather than another diagnosis app.

The deep research report's conclusion was unambiguous: build an "assisted decision funnel," not a chat-only system. Farmers should select from problem cards and be routed to specialist agents with clearly bounded scopes and hard guardrails — not be handed a blank prompt and left to phrase their own question.

The core decision: a guided funnel, not a chatbot

The single most consequential design decision was to reject the blank-chat interface. A chatbot demands that the user already knows how to ask — it front-loads typing, literacy, and the cognitive work of framing a problem onto someone who is in a field, possibly on a shared phone, under time pressure. For this audience that is the wrong default.

Instead, the Farmer Hub opens on a simple question — "What do you need today?" — and a grid of problem cards. The farmer taps a card; a short, picker-driven wizard collects only the data that card's specialist agent needs; the agent returns a structured, explained recommendation. This is the same template-first philosophy that drives the rest of EcoDesign's wizards, applied to the hardest-to-reach user: it removes the blank-canvas problem and replaces open typing with taps, photos, and choices.

It also makes the AI bounded. Each card maps to one specialist agent with a defined scope, defined inputs, and defined limits — rather than one general assistant trying to be a plant pathologist, hydrologist, and economist at once. Bounded scope is what makes guardrails enforceable, which matters enormously in agriculture where bad advice has a real cost.

Six modules, one hub

The Hub ships six modules. Each is mobile-first — the farmer is not at a desk — and each grounds its reasoning in the farm's own site data plus EcoDesign's curated component database.

1. Pest & Disease Triage

Photo-first, because a photo is faster and more reliable than a text description when literacy or typing is a constraint. The farmer uploads one or two photos; Kindwise crop.health performs the visual identification; an AI agent then combines that species ID with the farm's crop, growth stage, season, and recent weather to produce an IPM-first action checklist. Confidence is shown explicitly, and when the identification is inconclusive or the case is high-risk, the system escalates to "consult an agronomist" rather than inventing certainty.

2. Weather Alerts & Actions

Pulls extended Open-Meteo data and turns it into farm actions, not just numbers — frost warnings, heat-stress alerts, optimal sowing windows — contextualised to the farmer's specific crops and their current growth phase. This is the module that drives daily engagement, which the research predicted and which matters: a tool farmers open daily is a tool they trust when the harder decisions come.

3. Irrigation Guide

This one is deliberately not an LLM. The schedule is computed rule-based from reference evapotranspiration (ET₀, FAO-56 Penman-Monteith), a crop coefficient, and the crop's growth stage — ETc = ET₀ × Kc × stage factor — producing a seven-day irrigate / skip / check-soil plan. Water math should be deterministic and explainable, so it is. We use AI for language and reasoning, and algorithms for the things algorithms do better — the same principle that governs the rest of the platform.

4. Soil Health Check

Uses a SoilGrids ISRIC baseline (a 250 m modelled estimate, always labelled as an estimate, never as a lab result) plus the farmer's own observations to diagnose likely soil constraints, then recommends targeted amendments, cover crops, and practices from the component database with a realistic timeline.

5. Transition Coach

The differentiator. Conventional-to-regenerative is exactly the transition-first gap the research identified — and it is also where farmers are most afraid of losing income. The coach offers goals in plain farmer language across Sustainability and Livelihood categories, and a flexible duration (90 days, 6 months, or a full year). The plan, the check-ins, and the level of detail all scale to the chosen duration, and the language is risk-aware throughout: low-regret steps first, trial plots, don't change everything at once.

6. Crop Planning

Score-based suitability, not green-light / red-light flags. A crop is scored against the farm's real soil profile, climate, rainfall, drainage, and topography, with explicit tiers (Excellent / Good / Suitable / Worth trying) and the reasons behind the score shown. Honest scoring lets a farmer make trade-offs instead of trusting a verdict. When the best answer is a plant that isn't yet in the database, the AI says so and the farmer can request it — their suggestion shapes what we add next. Crop Planning unlocks only after Soil Health and Irrigation have been run, so the plan rests on a real data foundation rather than guesses.

Across all six, results are saved as session history and recommendations can be added to a personal Practice Library — so a farmer can revisit what was advised, track what they tried, and see what worked. Continuity is itself a trust mechanism.

The Farmer Hub: six modules, one mobile-first experience

Grounded in a curated component database

None of the modules answer from the model's general knowledge alone. Every agent receives a pre-filtered set of matched components from EcoDesign's curated databases — flora, fauna, soil amendments and substrates, technologies, farming systems, and garden types — selected for the farmer's specific site and problem before the agent ever runs.

That matters because it changes what the AI is doing. The pest module doesn't free-associate a treatment; it is handed the beneficial predators whose target matches the identified pest, the companion plants, and the amendments that suit the farm's conditions, and asked to assemble them into a coherent IPM response. The crop module scores real flora records against real site data. The transition coach builds its plan from practices and systems that actually exist in the database and fit the farmer's goals.

This is the same connection-oriented data model that underpins the rest of the platform — every component carries its inputs, outputs, services needed and provided, suitability conditions, synergies, and conflicts. The Farmer Hub is, in effect, that model made accessible from a phone in a field. We cover the schema and the systems thinking behind it in Designing for Resilience.

Guardrails: IPM-first and decision-support by design

In agriculture, an over-confident wrong answer has a real cost — a wasted spray, a lost crop, a poisoned beneficial population. The research was emphatic that safety has to be designed in from the start, not bolted on. So the Hub enforces a small set of hard rules:

• IPM-first. Plant-protection guidance defaults to biological and cultural controls. The system does not output pesticide dosages and does not recommend specific commercial products.
• Confidence and uncertainty are shown. A low-confidence identification is presented as low-confidence, not laundered into a clean answer.
• Human escalation. Inconclusive diagnoses and high-risk cases route to "consult a local agronomist" with a sharable summary, rather than a confident guess.
• Decision support, not prescription. Externally sourced data (SoilGrids soil baselines, Open-Meteo weather) is labelled as an estimate, and every result carries a verify-locally disclaimer. The Hub helps a farmer decide; it does not decide for them.

We treat the Farmer Hub as decision support, not a regulated agronomic prescription service. That is not a legal hedge bolted on at the end — it shaped the routing, the escalation triggers, and the language of every module from the first design session.

How the Hub answers each adoption barrier

It is worth being explicit about the mapping, because each design choice traces directly back to a documented barrier.

• Low digital literacy and time scarcity → problem cards and picker-driven wizards instead of blank chat; few taps to an answer.
• Photo is easier than text → photo-first pest triage as the flagship interaction.
• Scepticism toward "AI" advice → visible reasoning, explicit confidence levels, and score-based suitability you can interrogate rather than a verdict you must take on faith.
• Cost sensitivity and no hardware → core value works from a phone, location, and photos; no sensors required. Open data does the heavy lifting.
• Fear of transition risk → a transition coach built around low-regret steps, trial plots, and pacing, not an all-or-nothing switch.
• Field connectivity is unreliable → the modules need a connection for AI and image recognition, but every result, the Practice Library, and project data are stored server-side, so a farmer's history is available from any device the moment they reconnect. We were honest with ourselves that a true offline-first experience is a larger build than v1; server-side continuity is the pragmatic version of that promise.

The Hub also serves both of the smallholder audiences the research identified: conventional farmers who want to transition toward sustainability, and farmers already practising regenerative or agroecological methods who want to refine and de-risk what they do. The same six modules serve both — the Transition Coach simply meets each farmer at their current stage.

Where the Hub fits in EcoDesign

The Farmer Hub is the operational layer of EcoDesign — the day-to-day decisions that happen after, or alongside, whole-system design. A farmer can start here, solving immediate problems, and the data they accumulate (their soil checks, their crop plans, their practice library) becomes the foundation for fuller whole-farm design later. The relationship runs both ways: design produces the system; the Hub helps run it well, season after season.

It is also a concrete example of the platform's broader thesis — that AI is most valuable when it makes expert, systems-aware reasoning accessible to the people actually stewarding land, at a fraction of the conventional cost, with honesty about uncertainty built in. If you want the wider picture, start with our complete guide to permaculture design software, or see the Farmer Hub overview for small-scale farmers for the module-by-module and pricing detail.

Frequently asked questions

What is the EcoDesign Farmer Hub?

The Farmer Hub is a mobile-first decision-support tool for small-scale farmers. Instead of a blank chatbot, it presents problem cards that route the farmer to one of six specialist AI modules — Pest & Disease Triage, Weather Alerts, Irrigation, Soil Health, Transition Coach, and Crop Planning. Each module is grounded in the farmer's own site data and a curated ecological component database, and every recommendation is framed as decision support to verify locally, not a prescription.

Why did EcoDesign build a separate hub for small-scale farmers?

Small-scale farmers face heterogeneous, high-frequency problems but rarely have affordable access to soil scientists, agronomists, and irrigation engineers. Deep research into smallholder adoption showed the main risk is not model accuracy but adoption — connectivity, literacy, trust, cost, and time scarcity. The Farmer Hub was built as a guided, low-typing, guardrail-heavy funnel specifically to clear those barriers rather than expose the full design platform.

How does the Farmer Hub handle pest and disease problems safely?

The farmer photographs the affected plant. Kindwise crop.health identifies the likely species, then an AI agent combines that with the farm's climate, crop, and season to produce an Integrated Pest Management response — beneficial predators, companion plants, and cultural controls first. The system never outputs pesticide dosages or brand recommendations, shows confidence levels, and escalates to a human agronomist when the diagnosis is inconclusive or the case is high-risk.

Are the Farmer Hub's recommendations a substitute for an agronomist?

No. The Farmer Hub is explicitly decision support, not a regulated agronomic prescription service. Externally sourced data such as SoilGrids soil baselines and Open-Meteo weather are labelled as estimates, every result carries a verify-locally disclaimer, and high-risk or low-confidence cases trigger human escalation rather than confident advice.