Eggroll
Evolution-guided algorithm for hyperscale AI model training.
The founders
The company
Eggroll is built on EGGROLL, an evolution strategies algorithm for training large AI models without backpropagation. It creates model variations with low-rank perturbations in place of full-rank ones, which cuts the compute and memory that made evolution strategies impractical at scale.
The research behind it, published as Evolution Strategies at the Hyperscale around November 2025, reports up to a hundredfold increase in training throughput, reaching up to 91% of pure batch inference speed. Because it needs no gradients, it can optimise models in non-differentiable settings and pre-train language models that use only integer datatypes.
Eggroll is based in Oxford and was founded in 2026. The method has open-source implementations in JAX and C, and is aimed at researchers and developers training billion-parameter models.
Founder Juan Duque previously worked at Microsoft, LinkedIn and CreepShield. The paper's authors include Bidipta Sarkar, Mattie Fellows and Juan Agustin Duque.
Who else is building in this space
Companies working on the same thing, at every stage.
Unicorn
Enterprise-focused foundation models and AI systems
Scale-up
Empowers millions to build machine learning applications
Scale-up
AI workload management and GPU orchestration software
AI observability platform for model monitoring and explanation
Multimodal world models simulating reality for physical AI.
Human data platform for AI model development.
Thousands more founded every month.
Dealroom tracks new companies as they appear — founders, funding and signals, over the API or in the platform.