Skip to main content
Labvana
A Labvana productIn development

VanaNexus makes ML and AI usable in real R&D.

One place where experimental data becomes a workflow, a workflow becomes a model, and every result can be traced back to the run that produced it.

Watch the DemoRequest a Walkthrough

What you can build

Example areas teams are running today

Every screen below is a real workflow or AI agent, not an illustration.

Cheminformatics and molecular ML

Go from a SMILES column to a trained model without leaving the workflow. Descriptors and fingerprints feed classical models; message-passing graph networks and 3D representations take over when structure carries the signal.

A molecule descriptor node showing a rendered 2D structure beside a descriptor table listing molecular weight 457.432, exact molecular weight 457.158, and a heavy atom count of 32, each with a plain-language explanation
1 / 3
A molecule descriptor node showing a rendered 2D structure beside a descriptor table listing molecular weight 457.432, exact molecular weight 457.158, and a heavy atom count of 32, each with a plain-language explanation

Predictive modelling and evaluation

Train competing models against the same split and compare them honestly. Feature importance, cross-validated accuracy, and permutation scores are nodes in the workflow, so the comparison is recorded rather than reconstructed from memory.

A Gaussian process regressor node showing a feature importance table with standard deviations, the fitted kernel, and a training RMSE of 0.0626, on a canvas where gradient boosting and random forest models run alongside it
1 / 3
A Gaussian process regressor node showing a feature importance table with standard deviations, the fitted kernel, and a training RMSE of 0.0626, on a canvas where gradient boosting and random forest models run alongside it

Bayesian optimization and experiment design

A loop, not a one-shot fit. Teach a surrogate what you have already measured, ask it which experiments to run next, then feed the results back. Acquisition by expected improvement, probability of improvement, or upper confidence bound.

A Bayesian optimizer teach node showing an updated optimizer saved from a 10 by 9 training set, targeting concrete compressive strength, with a preview of the standardized cement, blast furnace slag, and fly ash columns
1 / 3
A Bayesian optimizer teach node showing an updated optimizer saved from a 10 by 9 training set, targeting concrete compressive strength, with a preview of the standardized cement, blast furnace slag, and fly ash columns

Quality, measurement, and statistics

The methods a process engineer already reports. Gage R&R across operator, part, and trial with AIAG interpretation; control charts that flag rule violations against control and specification limits.

A Gage R&R node reporting 28.75% study variation and 8.27% contribution against the AIAG acceptability guidelines, with a variance breakdown pie chart splitting operator variance and measurement variance
1 / 3
A Gage R&R node reporting 28.75% study variation and 8.27% contribution against the AIAG acceptability guidelines, with a variance breakdown pie chart splitting operator variance and measurement variance

These are a sample of what teams have built, not the boundary of what the platform does — the node library below goes considerably further.

Node library

Many analysis nodes, ready to connect

Not a framework you extend before it does anything — the methods R&D teams actually reach for, already built and configurable on the canvas.

Data preparation

Get real experimental data into a shape a model can use.

  • Load CSV
  • Filter rows
  • Standardize
  • Train/test split

Statistics and quality

The methods process and measurement teams already trust.

  • ANOVA
  • Gage R&R
  • SPC control charts

Machine learning

Fit, score, and reuse models against experimental data.

  • Gaussian process
  • Random forest
  • Gradient boosting
  • SVM classifier

Deep learning

Molecular and image models, trained or loaded from a checkpoint.

  • DMPNN
  • UniMol representation
  • ResNet-50

Cheminformatics and materials

Turn structures and formulations into model-ready features.

  • Molecular descriptors
  • Molecular fingerprints

Exploration and visualization

See the structure in the data before committing to a model.

  • PCA
  • t-SNE
  • Plots

Design of experiments

Close the loop: fit a surrogate, propose the next experiment, feed the result back.

  • Bayesian optimization
  • Optimizer teach
  • Optimizer query

Spectroscopy

Instrument output handled as a first-class input.

  • UV-Vis analysis

AI agents

Agents that sit on the canvas as nodes, reading the run's data and passing their work on.

  • AI agent
  • Document upload

Custom nodes

Where the platform stops guessing and you take over.

  • Python code

There is more in the library

These are a few of the nodes in each area. We walk through the full library with teams we work with — tell us what your analysis needs and we will show you what is already built for it.

Contact Labvana

Node creation

Build the node you are missing

The node library is not a fixed set. When the method you need is not there, you build it once as your own node — with an agent helping you write and test it — and from then on it sits on the canvas like every other node, and your agents can call it.

1 · CUSTOM CODEcustom_codesignalwindowresultdef smooth(signal, window): k = ones(window) / window out = convolve(signal, k) return out2 · SAVED AS A NODE</>smooth_signalSmooth Signal · custom nodeNode libraryData prepStatisticssmooth_signal3 · AGENT SKILL✦Agentruns in the workflowSKILLSsmooth_signal✓
  1. 1

    Write it once

    Start from a custom code node: declare its inputs, parameters, and outputs, then write the Python that does the work. An agent can draft and test it alongside you.

  2. 2

    Save it as a node

    Turn the working node into a template. It joins the node list under a name you choose, ready to drop into any workflow.

  3. 3

    Hand it to your agents

    Export it as an agent skill, and your agents can call it themselves as part of a run.

Who it is for

R&D teams without an ML group

No ML engineer needed to get a first result. Bring the data and the scientific question — the workflow, the model, and the record come with it.

Scientists who already model

Full parameter control and your own saved models, with the run history kept for you instead of by you.

Process and quality engineers

Gage R&R, SPC, and ANOVA in the same place as the models that use their output.

Working in

  • Materials science
  • Chemistry
  • Biopharmaceutical research
  • Manufacturing and process engineering

Work with us while it is being built

VanaNexus is in active development with a small group of design partners in materials, chemistry, and process engineering. Partners put it on real projects early and shape what gets built next.

Partner With UsVisit VanaNexus