KARAZ / Platform
The engineFive feeds in. Four decisions out.
Everything below exists to answer one question at a time: given what this block is actually doing this week, what is worth doing about it, and by when. The machinery is deliberately unglamorous — the difficulty in perennial crops is never the model, it is joining a soil probe, a satellite pass, a forecast and a price list into one sentence a grower can act on.
The feeds, and what each one is actually for.
Grower knowledge
The only feed that carries causation rather than correlation. Why the north block always cracks, which row floods, what happened the last time it hit 38 °C in May. It is captured as structured field observation, not free text on a clipboard, and it is weighted — a grower's judgement outranks a satellite index when the two disagree about a block they have farmed for twenty years.
Sensing at the block
Soil water content at multiple depths, canopy temperature as a stress proxy, air temperature, humidity, wind and solar radiation at the microclimate, and stem-diameter variation at sub-micron resolution — the plant telling you it is thirsty before the leaf does. LoRaWAN to a solar-powered cellular gateway, roughly every fifteen minutes.
Satellite
Sentinel-1 and Sentinel-2 vegetation and water indices at 10 m — NDVI, NDRE, NDWI — refreshed every few days. Sensors tell you what is happening at a point; satellite tells you whether that point represents the block or is the one anomaly in it. Neither is sufficient alone, which is the entire reason both are here.
Reanalysis and forecast
ERA5 reanalysis for the historical record the models train against, and operational forecast downscaled to block resolution for the window where action is still possible. A frost warning for the valley is not a frost warning for the block at 1,250 m on the north aspect — and that difference is usually the whole decision.
Arrival windows and price
Where the volume is expected to land, what the segments are paying, and when the regional glut opens. This is the feed almost no production platform carries, and it is what turns a harvest-date prediction into a commercial decision instead of a calendar entry.
Measured fruit and vine
Brix, firmness, calibre, colour, bunch and cluster counts, phenological stage on the BBCH scale, logged against the block that produced them. This is what closes the loop: every prediction the system makes is eventually scored against fruit somebody actually measured.
Hardware is specified per site and vendor-neutral — the platform ingests over LoRaWAN and MQTT rather than binding to one manufacturer's cloud. Where a site already has instrumentation, KARAZ reads it rather than replacing it.
Built to predict a change, not a state.
Forecasting an absolute value in a perennial orchard is easy to do and nearly useless — the season's level is dominated by the site, and a model that learns the site scores well while knowing nothing. KARAZ is delta-first: it predicts the movement from where the block is now, which is the only part a grower can influence.
| Stage | What it does | Why it is there |
|---|---|---|
| Feature extraction | Variational autoencoder compresses the raw multi-sensor stream into a small latent state per block. | Sensor channels are correlated and noisy. The latent state is what survives that. |
| Spatial integration | Lightweight graph network relates blocks at three spatial levels — tile, block, site. | Neighbouring blocks are not independent. Treating them as independent throws away the estate. |
| Temporal forecast | Temporal convolutional network projects the trajectory forward over the actionable horizon. | Decisions have deadlines. The horizon is set by how long the intervention takes to work. |
| Validation | Walk-forward on rolling multi-week windows, with a causal-feature audit on every input. | Prevents look-ahead bias and the leakage that makes offline scores flattering and useless. |
| Phenology | BBCH stage progression per variety per block, with confidence banded by data source. | Harvest date is a phenological question first and a market question second. |
Confidence is shown, not hidden.
Every projection is banded by what is driving it — measured archive, live forecast, or climate-average fallback. A grower reading a harvest window is told which of the three they are looking at, because a decision taken on a climate average deserves less conviction than one taken on a sixteen-day forecast.
Why there is no headline accuracy figure on this page.
Single-number accuracy claims for agricultural models are almost always measured on a setup that would not survive a leakage audit. Ours are re-audited under the causal-feature discipline described above, and we quote them per crop, per target and per site under NDA — not as a marketing number. If a supplier gives you one percentage for everything, ask them which target, which season, and which validation split.
Interoperable by construction, not by adapter.
An estate that adopts a closed platform inherits its ceiling. KARAZ is built on open standards so the data model survives us: the same context model an EU dataspace expects, the same transport an existing IoT estate already speaks.
| Layer | Standard | Role |
|---|---|---|
| Device ingest | MQTT via FIWARE IoT Agent | Sensor telemetry into the platform without vendor lock-in. |
| Context model | NGSI-LD, FIWARE Context Broker | The shared data model — estates, sites, blocks, varieties, observations, operations. |
| Application APIs | REST / HTTPS | Integration with farm-management, ERP and packhouse systems. |
| Edge-to-cloud | TLS 1.3, OpenZiti secure fabric | Zero-trust tunnels for sites with no public inbound exposure. |
| Federation | COP-PILOT Cluster 4 domain node | Service registration, secure inter-fabric connectivity, natural-language query over production context. |
Deployment options run from fully hosted to client-local, where processing happens on premise and only the outputs the grower chooses to share ever leave the site. Cooperatives and estates with sovereignty constraints get the second one — that architecture exists because a partner needed it, not because it reads well.
Model code and NGSI-LD schemas are published under an open licence and datasets deposited where the funding terms require it. The defensibility of KARAZ is the loop between agronomic decision and market result, not a hidden equation.
KARAZ.io — where the estate is written down.
The decision feed is the visible part. Underneath it is the thing most estates are genuinely missing: a single structured record of what is planted where, what was done to it, and what came off it — the record that makes both the model and the audit possible.
Estate → site → block → variety
Geometry drawn or imported from cadastral data, blocks split by variety, rootstock, planting year and training system. Everything the system says is addressed to a block, not to a farm.
Phenology, per variety
BBCH stage progression on one timeline per variety per site, with the predicted harvest window and the observations that corrected it — so you can see where the model was wrong and by how much.
What was actually done
Irrigation, fertigation, protection, thinning, covering — logged against block and date from the field, on a phone, in the time it takes to walk a row.
Decisions with deadlines
Exception-driven. Severity, the block, the reason, and the date after which acting no longer helps. Silence is a valid output and the system uses it.
Audit-ready by default
Because operations are logged against blocks with timestamps and provenance, the record that certification schemes ask for is a by-product of running the system rather than a fortnight of reconstruction in an office.
Block against block
The same variety at two altitudes, the treated split against the untreated one. This is how an estate learns which of its own beliefs are true — and it is the only honest route to the twenty-year variety decision.
From first call to first decision.
A KARAZ deployment is not a software rollout, it is a season. The sequence below is roughly how it goes; the pace is set by the crop calendar, not by us.
-
Scope the estate, honestly
Crop, varieties, altitudes, block geometry, what instrumentation already exists, and what the last three seasons actually cost. If the answer is that KARAZ would not change a decision you are already taking well, we say so on that call.
-
Write the estate down
Sites and blocks into KARAZ.io, with variety, rootstock, planting year and training system. Historical records imported where they exist. This is the least glamorous step and the one that determines everything downstream.
-
Instrument what matters
Sensing placed against the blocks that carry the decisions, not spread evenly for appearances. Satellite and climate feeds attach without hardware, so a site can begin producing useful output before a single probe is in the ground.
-
Calibrate against your fruit
Ground-truth measurement through the season — Brix, firmness, calibre, colour, stage. The first season is where the model learns your site; we are explicit that it is a calibration season and we price and plan it as one.
-
Run the loop
Decisions issued with deadlines, operations logged against them, outcomes measured. Every prediction is eventually scored against fruit somebody weighed.
-
Close it at the market
Predicted arrival and quality per block matched against the window where the premium is, early enough that contracts, pickers and cold chain can still be moved. This is the step that pays for the other five.
Shared infrastructure. Separate strategy.
An orchard and a vineyard both fail against weather and both sell into a market, and that is where the similarity stops. Roughly 60–70% of what a new crop needs already exists the day it is added. The remainder — the part that decides whether the system is any good — is written from scratch for that crop, with people who grow it.
| Component | Transfers | Note |
|---|---|---|
| Model pipeline | Fully | Same architecture, different target variables. |
| Sensing stack | Fully | Identical hardware and ingest path. |
| Satellite processing | Fully | Same Sentinel pipeline and indices. |
| Data architecture | Fully | NGSI-LD schemas extend per crop. |
| Phenological tracking | Concept | BBCH exists for each species, but the scale differs. |
| Quality model | Branch-specific | Interventionist for cherry, interpretive for vine. |
| Market bridge | Branch-specific | Market-window routing for fruit; style-to-segment matching for wine. |
| Fertigation programme | Branch-specific | Different crops, different regimes, different agronomists. |
| Post-harvest module | Cherry only | Wine is stable in tank; cherry has days. |
Cherry & stone fruit
Interventionist. Six steerable quality parameters plus the arrival date that prices them. Three altitudes in the Mediterranean, one cross-climate site in northern Germany.
Vine
Interpretive. Read which style the season is producing while it can still be changed, then match it to the demand segment that pays for it. Validated on a Tuscan estate.
Olive — next
Same pattern, own quality model and market logic. Not added before the two live branches carry their own weight. A platform that claims five crops and understands none of them is the failure mode we are most determined to avoid.
What KARAZ is not.
A short list, kept deliberately short. Any of these can be integrated alongside KARAZ; none of them is KARAZ, and we would rather say so here than in month four of a deployment.
- Not post-harvest logistics. Once the fruit is off the tree, the outcome is largely fixed. KARAZ works in the weeks before that, where it can still be changed.
- Not a monitoring dashboard. If the output of a season is a set of charts somebody has to interpret, the system has failed. The output is a dated decision or nothing.
- Not a substitute for breeding. Climate-resilient genetics take years and commit you for a decade once planted. KARAZ delivers resilience on the trees already in the ground, through management — complementary to breeding, not competing with it.
- Not autonomous. Nothing opens a valve on its own. Every decision is issued to a person who can overrule it, and the overrule is recorded and learned from.
- It is a decision system for the pre-harvest window, for crops you plant once and live with for decades — which is a narrow enough job to do properly.