Why we started
Agricultural advisory in Indonesia is almost entirely informal. The formal system, which includes Penyuluh Pertanian Lapangan (PPL) extension workers deployed by the government, is structurally underfunded relative to the number of farmers it is supposed to serve. The national ratio of extension workers to farm households has been discussed in Kementan policy documents for years. The practical consequence in West Java is that a PPL officer covers multiple kecamatan, visits individual farmer groups infrequently, and when they do visit, provides advice calibrated to the regional average rather than the specific plot.
The rest of the advisory landscape is informal: input dealers who have an incentive to sell product, farmer group meetings where the best-informed neighbour shares what worked for them, and occasional recommendations passed through WhatsApp group chats. None of this is bad per se. It reflects how information actually moves through rural networks. But it means that a farmer making a variety selection or a fertilizer dosing decision is working from a very wide range of information quality, and the quality correlates weakly with what is actually right for their plot.
We built Elevarm because we thought the information problem was tractable. Not all of it, and not all at once. But enough of it that a small focused product team could make a meaningful difference to the decisions a farmer makes in the four to six weeks before each planting season.
What the first year taught us about informality
The first thing we learned is that not all informality in agricultural practice is a bug. Some of it is accumulated local knowledge that outsiders systematically undervalue. Farmers in the Priangan highlands of West Java know which wet seasons have early-onset rain that pushes planting windows forward and which ones come late. They know which plots in the village drain slowly after heavy rain and which ones can hold moisture through a dry spell. That knowledge lives in their decisions implicitly. A digital advisor that ignores it produces recommendations that the farmer correctly identifies as incomplete and therefore discards entirely.
The practical consequence for our product design was that we had to build input mechanisms for local knowledge alongside the model-based recommendations. When a farmer registers a plot, we ask questions that elicit qualitative field descriptors: how long does water typically stand after a heavy rain, what is the highest yield this plot has produced in the last three seasons, what variety did you use when you had your best result? These are not rigorous soil science measurements. But they help the model parameterize the recommendation toward the farmer's actual field conditions rather than the soil class average.
What we got wrong about data collection
We initially designed the onboarding flow for literacy and smartphone comfort levels that do not describe our average farmer. We assumed farmers would fill in fields and enter numbers with reasonable accuracy. They can, but the cognitive load of a long form is a barrier to completion that we did not adequately test for before launch.
The simplified two-screen onboarding we use now collects far less data per farmer than the original version, but it gets completed at a much higher rate. We recover additional data through the season as farmers interact with recommendations and provide feedback, rather than trying to get everything upfront. This is a lesson we should have drawn from looking at how WhatsApp-based agri-services in Southeast Asia have been designed. The constraint is not what information the model needs. The constraint is what information a farmer is willing to provide in the time they have.
On trust and the first recommendation
Trust in a new advisory tool is built recommendation by recommendation. A farmer who follows the first recommendation and finds it was wrong, or confusing, or impossible to execute in their local context, will not follow the second one. This sounds obvious. It is less obvious how stringent the threshold is.
We treat the first recommendation as the most important product decision we make for any given farmer. It needs to be specific enough to be actionable, calibrated well enough that the outcome is better than what the farmer would have done by default, and small enough in scope that the farmer feels the risk of following it is manageable. A first recommendation to switch variety from the one they have used for six seasons to a different one they have never grown is not a good first recommendation, even if the soil-class analysis says it is objectively better. The second or third interaction is the right time for that conversation.
The parts of informality we are not replacing
We are not trying to replace the PPL system, the farmer group network, or the relationship between a farmer and their long-standing input dealer. Those relationships have functions that a digital platform cannot replicate. A PPL officer who physically walks a plot and notices signs of a water distribution problem with the irrigation bund is doing something our sensor-free model cannot do. A dealer who has been serving a village for 20 years and extends credit based on personal trust is providing something more than product fulfillment.
What we are trying to replace is the worst-case version of informal advice: the generic recommendation made without any knowledge of the specific plot, the product sold because the dealer had excess stock, the harvest timing decision made because it is what the farmer's grandfather always did at this time of year. That is the gap where specific plot-level information, applied consistently, makes a real difference. The first year showed us that the gap is real and that farmers value good information when it proves itself across a season.
Where we are at the end of the first year
We finished the first full planting season with early-access farmers across three West Java regencies. The platform works. The recommendations are technically sound and agronomically calibrated to West Java conditions. Farmers are returning for the second season, which is the most meaningful signal we have that the product is producing real value and not just novelty.
The harder problems are ahead: deeper variety coverage across the full diversity of heirloom and improved seed varieties grown in the Priangan highlands, better modelling of micro-climate variance across regency subzones, and expanding the Mitra dealer network into kecamatan that currently have no digital connectivity at the dealer level. None of that is novel as a product roadmap. What is specific to us is that we are trying to do it at a price point and interaction design that works for farmers who do not own a laptop and whose primary connection to the digital economy is a mid-range Android phone and a WhatsApp account.