Case 02 · AgroTech · 2026
Livestock-management SaaS — MVP with a pilot running. Farm ERP, digital corral, NDVI via Sentinel-2 and an AI consultant with tool calling integrated with Postgres.
Context
01 · Context
Brazil's livestock management market is dominated by ERPs with UX from the 2000s — heavy desktop apps, no AI, no satellite vision. Independent livestock consultants charge R$ 3-8k/month (USD ~600-1.5k) and still operate reactively — they show up after the cattle already lost weight.
02 · Problem
Ranchers decide pasture rotation, supplementation and AI breeding with days or weeks of delay. Late decisions in livestock are absolute loss — kilograms of meat that didn't gain. Anyone running 2,000+ head loses thousands of dollars a month on decisions that could have been automated.
03 · Response
Multi-tenant SaaS combining farm ERP, voice + RFID corral, NDVI from Sentinel-2 satellite imagery and an AI consultant doing tool calling over real farm data. Positioning: the Linear of livestock — a space where no national player ever shipped contextual AI integrated with satellite vision.
Stack
No fashion picks. Each dependency here passed through a filter of maintenance cost, time-to-market and fit with a small team.
RSC for chart-heavy dashboards, Server Actions for mutations, Edge Runtime for AI routes
useOptimistic so the corral works offline-first without UI flicker
Multi-tenancy via RLS, realtime dashboards, storage for handling photos. 30+ tables prefixed agrotrack_
Tool calling for the consultant, ephemeral prompt caching in 2 layers, per-feature model routing
10m/pixel NDVI updated every 5 days, public free data — zero recurring acquisition cost
GeoJSON pasture polygons, NDVI raster overlay, client-side area and centroid math
Weight evolution, ADG by lot, cash flow, stocking rate — all composable and typed
PWA → native iOS/Android with 95% code reuse. Small team can't maintain 3 stacks
Architecture
Edge Functions handle heavy cron jobs (daily insights, NDVI refresh) and the AI chat route. Postgres is source of truth with farm_id RLS. Sentinel-2 enters through an Edge Function that normalizes tiles and stores aggregated results per polygon.
01 · client
PWA Next.js
iOS / Android via Capacitor
client02 · edge
Edge Functions
AI chat · Cron · NDVI sync
edgeApp Router RSC
Dashboard · Corral · Pastures
edge03 · data
Postgres + RLS
30+ tables, multi-tenant
dataStorage
Photos · KMZ · Reports
dataSentinel-2 ESA
NDVI 10m/5d
externalRFID wand
Bluetooth · Web BLE
externalLLM provider
Multi-tier routing
aiMain flows
Technical decisions
Each one carries the problem, the options considered, the choice I made and the price I paid for it.
Decision 01
Problem
General chat doesn't need the most expensive tier; NDVI analysis does. Using the most expensive model for everything explodes cost; using the cheapest sacrifices quality on critical decisions.
Options considered
Choice
An agrotrack_ai_config table per workspace maps feature → tier. Fast tier for chat and voice, balanced tier for NDVI insights and data crosses, premium tier for yearly strategy. Enterprise customers can promote all routes to the balanced tier via override. Provider-agnostic — swapping providers means updating the mapping only.
Tradeoff accepted
More config complexity and an extra lookup per call. In exchange: ~70% token cost reduction vs default top-tier, without sacrificing quality on important decisions.
Decision 02
Problem
Farm context (herd, pastures, current alerts) rarely changes between messages. Re-sending it on every call blows up input cost and latency.
Options considered
Choice
System split into BASE_SYSTEM_PROMPT (static, cache 1) + farmContext (semi-static, cache 2). Both with cache_control: { type: 'ephemeral' }. Re-warm happens transparently when context changes.
Tradeoff accepted
Cache expires in 5 minutes — sparse conversations pay the re-warm. In exchange: ~85% cache hit rate after stabilization and ~80% input cost reduction vs inline system.
Decision 03
Problem
Mid-size ranchers (1,000-3,000 ha) don't have budget for distributed IoT. Owned drones add equipment cost, maintenance and aviation regulation. Without pasture vision, the AI is blind.
Options considered
Choice
Sentinel-2 through an Edge Function that downloads tiles, crops them by the pasture's GeoJSON polygon, computes mean NDVI + 90-day trend, stores the aggregate. Raw tiles aren't kept — only the per-pasture aggregate.
Tradeoff accepted
Medium resolution can't see fences or troughs. In exchange: zero hardware, zero maintenance, zero recurring cost. The customer perceives value from day one.
Decision 04
Problem
Multi-farm SaaS with strict data isolation between customers. Schema decision affects complexity across the entire codebase — choice has to be deliberate.
Options considered
Choice
All tables have farm_id, RLS policies read from agrotrack_farm_members by auth.uid(). Three roles: owner, consultant (accesses multiple farms), employee (only their farm). Service role bypass only for internal Edge Functions.
Tradeoff accepted
Every app query carries an extra implicit filter. In exchange: 1 schema, 1 database, 1 deploy. Operationally irrelevant up to 10k tenants — beyond that there's room to partition by region.
Decision 05
Problem
The AI consultant has to answer about *this* farm, with fresh numbers. RAG over a daily dump hallucinates values. An answer like 'your pregnancy rate is at 45%' must come from the database, not from the LLM's imagination.
Options considered
Choice
Tools exposed to the agent: get_animal(tag), get_lot_performance(lot_id, period), get_pasture_ndvi(name, days), get_financial_summary(period), query_pasture_capacity(), schedule_activity(...). Each tool runs a parametrized query with RLS applied.
Tradeoff accepted
Every tool requires its own code and maintenance when the schema changes. In exchange: exact answers backed by fresh data, zero numeric hallucination, and the agent becomes an actual agent — not a decorated chatbot.
Decision 06
Problem
Ranchers are reactive — they only discover a problem (low ADG, degraded NDVI, empty cows post-AI) when they open the app. Without proactivity, the AI's edge dissolves after one week of usage.
Options considered
Choice
pg_cron triggers a daily Edge Function at 5 AM. It crosses lot ADG × pasture NDVI × monthly cost × days post-AI and creates 'Attention' cards in the ai_insights table + Web Push notification.
Tradeoff accepted
Edge Functions cap execution at 50s — giant farms (>5k head) require pagination. In exchange: zero extra infrastructure, fires every time the cron runs, costs a few cents per month.
Decision 07
Problem
The app has to run in the corral — bad signal, dirty hands, nervous cattle. It must install on the phone, work offline, sync when back at the office. A small team can't maintain iOS, Android and Web separately.
Options considered
Choice
PWA-first with Service Worker for offline and Background Sync for the operation queue. Capacitor packages the same build for App Store and Google Play, with native plugins only where strictly needed (camera, Bluetooth BLE for RFID).
Tradeoff accepted
Performance ~5-10% below pure native on highly complex transitions — irrelevant for the real use of this app. In exchange: 95% code reuse, 5x faster time-to-market, maintenance in a single tree.
Highlighted implementation
// AI Consultant: per-feature model routing + 2-layer prompt caching
async function chatWithFarm({
farmId,
message,
feature,
}: ChatRequest) {
const [config, farmContext] = await Promise.all([
getFarmAIConfig(farmId),
getFarmContext(farmId), // herd + pastures + alerts
]);
const model =
config.feature_models[feature] ?? DEFAULT_MODELS[feature];
return anthropic.messages.create({
model,
max_tokens: 1024,
system: [
{
type: "text",
text: BASE_SYSTEM_PROMPT,
cache_control: { type: "ephemeral" },
},
{
type: "text",
text: farmContext,
cache_control: { type: "ephemeral" },
},
],
tools: FARM_TOOLS, // get_animal, get_lot_performance, get_pasture_ndvi…
messages: [{ role: "user", content: message }],
});
}Technical metrics
Status
MVP
pilot running, pre-paying customers
Avg. cost per AI conversation
~$0.02
measured in pilot · 9 initial conversations
Satellite resolution
10m/pixel
Sentinel-2 ESA, 5-day revisit
Postgres tables
30+
100% under multi-tenant RLS
Tools available to the agent
6
animal, lot, NDVI, finances, capacity, schedule
Multi-tenant
RLS
farm_id isolation in every query
Retrospective
Decisions I'd reverse knowing what I know now. Not regret — the kind of learning that changes the next project.
01
I created tables (sync_queue, ai_insights, soil_analyses, pasture_managements) before having any code writing to them. Rich, empty schemas turn into cognitive noise. Now I only create a table when there's a seed migration + writer code ready.
02
The initial instinct was to use the balanced tier for everything — uniform quality. Mapping feature → tier added some extra complexity, but saves ~70% in token cost as measured in the pilot. Explicit per-workspace configuration also opens the door for a future Enterprise customer who wants to override defaults.
03
I considered RAG over a daily tenant dump. In practice, RAG hallucinates stale numbers — a rancher asks 'what's my pregnancy rate now?' and the LLM answers based on yesterday's snapshot. Typed tools querying Postgres in real time was more work but eliminated numeric hallucination.
Gallery
07 frames