AI research · for teams building world models
Your world model’s research team.
Ecological forecasting, remote sensing, species distribution. We took a 25M-parameter model past the 800M original on a public benchmark suite, shipped as pull requests.
+17.3% harmonic mean · 96.8% fewer params
We only get paid when your benchmark gets beat.
Automatic results
Breakthroughs in months days.
Where we work
breakthroughs in the field.
Climate, ecology& geospatial
Improve models for a changing physical world, without sacrificing the constraints that make results scientifically useful.
Medicalimaging
Improve diagnostic and imaging models with evidence your technical and clinical teams can review.
Scientific &industrial vision
Turn complex sensor and visual data into a model advantage: defect detection, forecasting, uptime.
Real-worldrobotics
Improve the models that make machines work beyond curated benchmarks.
In progress. Sign up for updates as we make these breakthroughs public.
Working in another technically difficult ML domain? If your model has a measurable objective and meaningful constraints,
From a customer
Ensue is a critical partner, helping us push the edge of what’s possible. Their team brings real respect for the depth of the science and consistently turns complex model iteration into results that are clear, credible, and commercially valuable.Read the DeepEarth case study →
Lance Legel
CEO · Ecological Intelligence, Inc.
Visiting Scientist, UC Berkeley · Former NASA
What you get
Science that scales.
01 · Model science
A better model than the one you have.
The loop runs thousands of experiments across architecture search, training schedules, and data curation, and converges on the combination a lone researcher would take six months to stumble into.
On one engagement, that meant beating an 800M parameter model with 25M parameters. Science over scaling, verified against the customer’s eval.
25M > 800Mparams, on the customer’s eval
02 · Inference and kernels
Faster on the hardware you ship on.
A model that’s slow on your hardware isn’t shipped. We search kernels, quantization schemes, and compilation targets, and land the speedup as a PR you can merge the same day.
We’ve done it on Apple silicon and Nvidia GPUs, across six generations of hardware, with the results published.
6.3×on Apple Neural Engine
Read the case study →Who you’re working with
“Your data becomes a moat the day it becomes a model. Your model.”
Austin Baggio
CEO · Co-founder
Your first call and your weekly report are with me.
Sai Vegasena
CTO · Co-founder
I design the evals and steer the loop.
“A single researcher tunes one thing at a time. We try fourteen approaches before lunch.”
Set the eval. Set the GPU budget. We only get paid when the number goes up.