UrbanKisaan (YC W20) — Foundation models and autonomous chambers for plant biology

In their words

UrbanKisaan is an AI-for-plant-biology company. We build foundation models and the physical infrastructure that trains them. Shipping a single new seed variety takes about ten+ years. Two on germplasm, three on crosses and selection, three or four on field trials, one on regulatory. Inside that decade, prediction and experiment never properly meet. DNA is abundant, phenotype data is scarce and unpaired, and the cycle never trains itself. We closed that loop. Three foundation models cover the upstream science. Aadi is a 48-layer transformer trained from scratch on the public and proprietary corpus of plant genomes; it reads plant DNA and ranks variants before any wet-lab spend. Jeeva is the genotype-to-phenotype model that turns a seed and a target environment into yield, flowering, maturity, and height with confidence intervals plus per-locus attribution. Akash is the geospatial backbone - it maps every pixel of every field to crop class and growing condition. Two pieces of physical infrastructure close the downstream half. GeneRush is a controlled-environment chamber engineered for speed: extended photoperiod, optimized temperature, integrated phenomics. Wheat reaches anthesis in 38 days. Breeding programs run 4–6 generations a year instead of 1–2. GrowRush takes that same chamber to commercial production scale, where every harvest is also a labelled phenotype training pair flowing back to Jeeva. The unit is paid for by what it grows; the science works because the unit is identical to the chamber the AI was trained on. Every cycle produces proprietary genotype-to-phenotype data the next generation of models needs. The ten-year breeding cycle compresses toward three. We are already commercial

What the record says