MANIFOLD
WERKS


mnfldwrks — models of the real world

01 — What We Build

World models
from data.

We build AI systems that learn mathematical models of biological systems from data. Not chatbots. Not pipelines. Models that simulate, predict, and learn from experiments.

Our models live on manifolds — the geometric surfaces that describe how cells move, differentiate, and respond. We learn the dynamics on these surfaces directly from high-dimensional biological data, then use those dynamics to do science.

Data
Geometric Embedding
Neural ODE
Simulation
Hypothesis
Lab
Repeat
02 — Flagship

GEOMANCER

Geomancer is an AI scientist. It takes biological data — single-cell transcriptomics, imaging, time series — and constructs a mathematical world model: a system of differential equations on a learned geometric space that captures how the biology actually works.

Then it uses that model. It runs simulations, generates testable hypotheses, designs experiments, and updates itself when the lab results come back. Geomancer learns ODEs from your data and tells you what happens next.

Geometric Data Encoding
Learns the intrinsic manifold structure of high-dimensional biological data. Preserves topology and geometry that linear methods destroy.
Neural Dynamical Modeling
Fits neural ODEs and SDEs directly on learned manifolds. Captures continuous-time dynamics without discretization artifacts.
Cross-Domain Generalizability
Same mathematical framework applies to neurodegeneration, cancer, stem cell biology. Geometry doesn't care about the domain.
Lab-in-the-Loop
Generates hypotheses, designs validation experiments, incorporates results. Closes the gap between computation and wet lab.
03 — The Loop

Closed-loop
discovery.

The core innovation isn't any single model. It's the loop: a system that learns, predicts, tests, and refines — continuously. Each pass through the cycle makes the model sharper and the hypotheses more precise.

04 — Applications

Where it
works.

Neurodegeneration
Modeling dopaminergic neuron dynamics in Parkinson's disease. Predicting disease trajectories and identifying intervention windows.
Cancer Biology
Simulating breast cancer organoid growth and drug response. Building predictive models of tumor evolution from single-cell data.
Stem Cell Biology
Mapping differentiation landscapes and predicting cell fate decisions. Guiding reprogramming strategies through learned dynamics.
05 — About

Built to
last.

Institutions
Yale University
 
Mila — Quebec AI Institute