We build pipelines,
not chatbots

What already works

A Telegram bot in Uzbek. Every morning at 06:00 Tashkent time: the day to irrigate and the cubic metres — and, on a pump, its running hours. The field is traced with a finger in the Mini App or uploaded as a boundary file, up to 10 fields per chat. After irrigating, the farmer taps "Suv berdim", and the next calculation starts from that fact.

Under the hood: an FAO-56 water balance on the weather forecast and Sentinel-2 imagery with Landsat 8/9 as backup — NDVI shows greenness, NDMI the water in the vegetation. The engine models drip with its wetted fraction. Without it, it told the pilot orchard to irrigate every 20-plus days while the partner farmer watered every four. We found the error by checking the advice against what the farmer does by hand — now it comes out at 4–5 days.

In closed demo

Weather-based disease risk, alfalfa cut tracking, a web app with an Android APK, and AI Agronom: chat and leaf-photo analysis that take the field's own calculation into account. We open them gradually: a raw feature must not reach a farmer before it has been checked.

Who buys

Farmers with a pump and drip: they pay for every hour the pump runs. Large farms and clusters — their fields are loaded from a boundary file, not one by one. For gravity-fed fields the payer should not be the farmer but districts, water user associations and donors: they need independent satellite accounting of water.

The stage, honestly

The first contract is signed: 100 ha with a large farm, billed monthly. The pilot runs on two fields of a partner farmer near Samarkand (≈4 ha), with the advice log kept since 9 August 2026. No confirmed water savings yet: the pilot fields have no meter, and we do not promise percentages until a measurement shows them. The calculation does not yet account for groundwater contribution or salt in the irrigation water.

Where the work is here

The main thing is not the model, it is merging duplicates: one apartment listed three times is one row, not three. The price is calculated in dollars: UZS has been moving, and the model would have learned the exchange rate instead of the price. Testing is on a later period only: a random split gives pretty but false numbers.

Three separate models, and the error is published for each one, not just the best: ≈ 15% for new builds, ≈ 25% for houses. Every night at 03:20 Tashkent time the pipeline collects listings and refreshes the market showcase on its own; a new model reaches the product only after passing quality gates.

The refusal rule

A district with fewer than 100 observations a year gets no number at all — no wide range, no "approximately". The threshold held even when the database grew to the whole country: an honest refusal beats a number with no market behind it.

The stage, honestly

The bot, the website and the market showcase run in production, but there has been no pilot with live clients yet, and nobody pays. The market made 171 thousand deals in the first half of 2026 without a single public deal price — so a realtor's first month is free in exchange for marking real deal prices.