In agriculture, a forecast without measured error is an opinion with a chart. We build on the data the farm already produces and validate every prediction against what actually happened last season.
4–6 weeks to production · 10 years of software engineering · 1 week to review whether the data is enough
Built on
Ninety seconds: who we are, how we work and what you get at the end.
Contracts, price lists, email, the CRM: what you already have becomes sourced answers and agents that do the work. Pick a case and follow the path.
Example: a field needs treating. The system works out the dose and schedules the job.
Connected by API or uploaded. Nothing leaves your infrastructure when the case demands it.
120 pp → 760 chunks
Searches by meaning and by exact word, and filters by the tag before choosing.
answers with source · 91 of 100
Swipe to follow the flow →
01
By the time damage is visible from the field, the loss has occurred. The intervention window was days earlier.
02
Decided on the experience of whoever has been there longest — far better than nothing, and not reproducible.
03
You handle what already happened because there is no way to see what is coming.
04
The operation grows and the capacity to observe directly does not.
05
The plot is not homogeneous, but it gets treated as if it were, because measuring the difference costs.
01
The entry condition
This is the sector where the most gets promised and the least can be sustained without a base. A model deciding irrigation or raising a pest alert needs history from that farm: what was planted, when, what was applied, what the weather was, what happened. Without it, what you can deliver is a regional average presented as a specific recommendation — exactly the kind of unbacked signal this site avoids.
So the first job is often not a model but the record: turning notebooks and loose sheets into queryable data, captured from a phone in the field rather than transcribed at night. It is not glamorous and it is the difference between a system that learns and one that guesses.
And validation: every prediction is run over last season and compared with what actually happened. An error comes out, and that error is what gets handed over. A forecast presented without error measured against history is a nice chart, and in a planting decision that costs the season.
The prediction run over last season, compared with the outcome that actually occurred.
What happens if it fails
We declare the model is not ready and keep recording data. We do not hand over a chart without an error figure.
How many days earlier it was detected, compared with when a person in the field would have seen it.
What happens if it fails
If the lead time is hours, it changes no decision and does not justify the cost.
Alerts that corresponded to nothing. Measured separately because it is what destroys trust.
What happens if it fails
The threshold goes up. A system that over-alerts gets switched off in two weeks and then helps nobody.
What share of field operations actually got recorded.
What happens if it fails
Without records no model is possible, so this number governs all the others.
If the farm has no historical records and no sensors, no model is possible and we say so on the first call. The right project there is the record — capturing well what is already being done — and talking about prediction one season later, with your own data.
And we do not sell computer vision over drone imagery without a labelled set from that region and that crop. A model trained on images from another latitude fails precisely on the local pests, which are the ones that matter.
Chunking, reranking and hybrid search, evaluated with recall@k and NDCG.
MCP servers against the systems already in operation.
Applied where there is volume and stable rules, not where there is expectation.
Architecture, evaluation criteria and cost per query before writing code.
Key takeaways
There is one entry condition here and it is not negotiable: without historical farm data there is no model. With one season of records you can start; with none, you cannot.
Every forecast gets validated against what happened: run it over last season and compare against the real outcome. That error is the number handed over, not a trend chart.
What pays most is not yield prediction but early warning: detecting sooner what today gets detected when it is already visible, which is when the loss has occurred.
Traceability is the least glamorous case and the clearest return, because the export buyer requires it and today it lives in notebooks and spreadsheets.
An AI system is worth what its sources are worth. These are the standard connectors; anything with an API or a database connects the same way, and what has no API is handled by file.
SAP
Enterprise ERP
Oracle
ERP and database
NetSuite
Cloud ERP
Salesforce
CRM and service
HubSpot
CRM and marketing
PostgreSQL
Database and pgvector
Microsoft SQL
Database
Snowflake
Data warehouse
BigQuery
Google data warehouse
Databricks
Data platform
Redshift
AWS data warehouse
Synapse
Azure data warehouse
Supabase
Managed Postgres
Workday
Payroll and HR
QuickBooks
Accounting
Sage
Accounting and ERP
Xero
Cloud accounting
Shopify
Catalogue and orders
WooCommerce
Catalogue and orders
Magento
Catalogue and orders
Stripe
Payments and subscriptions
Google Drive
Documents and folders
CSV y Excel
Flat files
Nothing in this category
We built and operated the assistant for a loyalty platform serving more than two million active users across ten countries.
We built the full pipeline: document ingestion and normalization, chunking, embedding generation, vector store on Azure AI Search and Pinecone, and retrieval with grounded generation on LangChain.
We held it above 99% availability for three years.
RAG systems →Historical farm records: what was planted, when, what was applied, what yield followed. With one complete season of orderly data a model can start being validated; with less, what can be delivered is a dashboard that organises what already exists and a good capture system, which is usually the first real project. Sensors and a weather station help a lot but are not the initial requirement: the record is.
Because it gets run over last season and compared with what actually happened, and that error is handed over before anyone makes a decision with the system. It is the same requirement we apply to any AI project, except here the consequence is a planting. Any forecast presented without error measured against history is a trend chart wearing the language of certainty.
It can be built, on one condition almost nobody states: you need a labelled image set from that region and that crop. A model trained on images from another latitude recognises what it saw in training and fails precisely on the local pests, which are the ones that matter. If that set does not exist, the first job is building it, and that takes a season.
It works by hectare and crop value, not by company size. A high-value crop with technified irrigation on fifty hectares has a better case than an extensive low-margin one on five hundred. And for a mid-sized farm the clearest return is usually traceability rather than prediction, because the export buyer requires it and today it lives in notebooks.
Yes, to weather stations, moisture sensors and irrigation systems that expose an interface, and to the farm management platforms already in use. When machinery exposes nothing, the system works with manual capture from a phone in the field, which done properly is worth far more than badly installed sensors nobody calibrates.
A recording and dashboard system, four to six weeks. A predictive model depends on the data: if an orderly season exists, another four to six weeks including validation; if it does not, recording comes first. It is quoted with fixed scope, price and date in a proposal 48 hours after the first call, and the review of whether the data is sufficient happens first, in one week.
It gets fixed, it gets built, or it is not worth it. And if the diagnostic does not reach three actionable findings, it is not charged.
We use cookies to improve the user experience. Privacy
