UrbanKisaanEnquire now

HomeIntelligenceJeeva

JeevaGenotype × environment model

Predict the field trial before you plant it.

Give it a genotype, a place and a planting date. Jeeva predicts yield, flowering, height and more, each with a confidence range. Licensed as weights you run yourself.

See the benchmarks
Illustration of Jeeva predictions for three lines across temperaturesPredicted yield for three made-up lines as the mean temperature around flowering rises from 24 to 36 degrees. Line A leads in cool seasons, Line B is the most stable, and Line C leads in the hottest seasons. The shaded ranges widen above 33 degrees, where fewer training sites exist.Fewer trainingsites this hot2462427303336Yield, t/haMean temperature around flowering, °CLine ALine BLine C
Ranking at this temperature90% range
  1. 1Line A5.3 t/ha5.0 to 5.6
  2. 2Line B4.9 t/ha4.6 to 5.3
  3. 3Line C4.5 t/ha4.1 to 4.8

Illustration with three made-up lines. The best line changes with the season.

580kfield trial records in training
~10countries of trials
6crops, from rice to chickpea
AnySNP panel, any density

How it is different

Built for how genes and weather interact.

  • Any marker panel

    Chip, GBS or sequencing, at any density

  • The field, described

    18 months of weather, satellite and soil data

  • Weather shapes the reading

    The environment changes how the genome is read

  • Every trait in one pass

    Related traits inform each other

  • Honest ranges

    A 90% range on every prediction

  • Shows its working

    A genome map behind each prediction

Used for

Multi-environment screeningClimate-resilient selectionFaster breeding cyclesFuture-climate stress testsTrial planningWhat-if simulation

Benchmarks

Tested on fields it had never seen.

Each test holds out a site and season the model never saw. Score: correlation between predicted and observed (Pearson r).

JeevaClassical model (G×E-GBLUP, or best baseline for maize)

Rice · 8 held-out environments

Ahead of G×E-GBLUP on all four core traits

Yield
0.484
0.429
+0.055
Days to flowering
0.815
0.813
+0.002
Days to maturity
0.815
0.812
+0.003
Plant height
0.759
0.748
+0.011

Rice yield · environment by environment

Higher in 7 of 8, level in the eighth

Yield accuracy in each of 8 held-out rice environmentsFor each held-out environment, a line joins the G×E-GBLUP score on the left to the Jeeva score on the right. Seven lines rise. One, from 0.546 to 0.535, is level.0.20.30.40.50.6Held-out environment 1: G×E-GBLUP 0.272, Jeeva 0.353Held-out environment 2: G×E-GBLUP 0.546, Jeeva 0.535Held-out environment 3: G×E-GBLUP 0.218, Jeeva 0.282Held-out environment 4: G×E-GBLUP 0.511, Jeeva 0.561Held-out environment 5: G×E-GBLUP 0.536, Jeeva 0.627Held-out environment 6: G×E-GBLUP 0.334, Jeeva 0.364Held-out environment 7: G×E-GBLUP 0.416, Jeeva 0.495Held-out environment 8: G×E-GBLUP 0.314, Jeeva 0.376levelG×E-GBLUPJeeva

Maize · 28 held-out environments

Biggest gains on traits driven by weather

Grain moisture
0.718
0.561
+0.157
Silking
0.573
0.383
+0.190
Anthesis
0.563
0.421
+0.142
Ear height
0.436
0.321
+0.115
Grain yield
0.426
0.386
+0.040

Across all 12 maize traits: ahead on 5, level on 3, behind on 4. Full table in the evaluation pack.

Rice · confidence ranges

Ranges you can plan with

  • Yield94.3%
  • Days to flowering92.7%
  • Days to maturity94.3%
  • Plant height89.5%

Share of results inside the 90% range. Orange mark: the 90% target.

Benchmark models are trained on benchmark data only, never on our private trial records.

Model card summary

Jeeva at a glance.

In

GenotypesVCF · HapMap · PLINK
Location and planting dateWeather, satellite and soil built in
JeevaRuns on CPU or GPU

Out

Trait values with 90% ranges
Genome map
Model size
Tens of millions of parameters
Delivered as
Weights and inference code
Licence
Research or commercial
Model card
Under NDA

How to get it

The model itself, on your own servers.

No platform and no uploads. Your data never leaves your network.

  1. Tell us your crop
  2. Sign an NDA
  3. Test it on your past trials
  4. License the weights

You receive

Model weightsInference packageModel cardEvaluation packSet-up help
Research licenceCommercial licence

Questions

Before you license it.

Is Jeeva an app or an API?

Neither. You license the trained model and run it on your own servers. There is no upload portal and nothing passes through us.

What hardware does it need?

It runs on an ordinary CPU server. A GPU makes large batches faster but is not required.

Will it work with our genotyping platform?

Yes. Jeeva reads each marker by its position on the genome, so chip, GBS and sequencing data all work at any density. Bring VCF, HapMap or PLINK files.

Our crop is not on the list. Can we still use it?

Tell us the crop. We can test how well Jeeva transfers to it and talk about tuning it on your own trials.

Can we check it before we buy?

Yes. Under NDA you get the full evaluation pack, and we run Jeeva on past trials you choose so you can compare it with what really happened.

What do you share about how it works?

Inputs, outputs, every benchmark and the model card. The architecture details, training recipe and private training records stay with us.

How is it priced?

It depends on your organisation and how you will use it. Research and commercial licences are priced differently. Ask us for a quote.

Works with Aadi

Jeeva predicts the field. Aadi finds the genes behind it.

See Aadi