I
Pipelines
Ingestion, exploratory analysis, statistics, feature engineering, model training and tuning as steps on one canvas.
Python and R, with no environment to configure.
See analyses in practice
A research bench for your whole lab
Where data
becomes discovery.
Build analyses in Python and R, keep every run reproducible, and bring your team’s findings to publication.
For universities and research laboratories. Hosted in Europe.
Visual pipelinesRecorded runsInstitutional controls
The five pillars
What the platform stands on
From the first dataset to the published finding, five commitments hold the work together.
II
Reproducibility
Versioned pipelines, pinned code, and tracked runs. Rebuild any result, any time.
Every agent action and data touch lands in an append-only audit log.
Follow the research recordIII
Sovereignty
Hosted in Europe. Your datasets and results are yours alone.
EU data residency. Your research data is not used to train third-party AI models.
See where data livesIV
Collaboration
Fourteen research roles, from department head to visiting researcher, with supervisor-approved onboarding.
Per-feature read and edit access, and every artefact inside a project.
Explore roles and accessV
Publication
Reports derived from the run itself, versioned and human-editable, with one-click PDF export.
Findings link across datasets and papers in the knowledge graph.
Read published researchThe people and the work
Run your lab like a lab
Organise the work around your team, from the department head to the visiting researcher.
- PeopleInvite researchers and set their roles and access.
- ProjectsKeep the data, pipelines and results of a study together.
- ReviewFollow progress and approve the actions that need a second person.
- DeliveryKeep versioned reports, models and publications with the work.
See how roles and delivery connect
AlchemYst-Crucible
Analysis, with review built in
The research question stays yours. The platform helps turn it into an analysis you can inspect and rebuild.
- PlanRead the data and design the analysis.
- WriteGenerate the Python or R steps.
- ReviewCheck the program before it runs.
- RecordKeep the code, parameters, figures and results.
Explore the analysis architecture
The intelligence layer
The agentic bench
One architecture that designs the analysis, reads your data, writes the program, checks it, and corrects it when it fails: reasoning where it earns its place, tested code where it does not, and a review before any of it runs.
AlchemYst-Crucible
On every planThe reasoning agent behind every pipeline
Thinks the analysis through before writing it, works the harder parts in parallel, then assembles and reviews the result against your data before it is allowed to run.
Every pipeline, every language.
The safety path is the floor, not the ceiling: it exists so the work continues when the reasoning layer cannot be reached. It all runs on European infrastructure and does not train on your data, and an institution can point it at its own inference account.
Who is on the bench
Live totals for research workspaces, projects, researchers and publicly shared articles.
9Research workspaces
14Research projects
10Researchers
8Publications shared
Key product features
The instruments
Every stage of a study, from the raw table to the published finding, on one bench.
Data, from wherever it is
Upload a file or an archive, query a database, call a REST API, or connect your institution's own storage. Columns, types and quality are read on arrival, so the first thing you see is what you actually have.
Exploratory analysis, unprompted
Distributions, correlations, missing values and a quality score the moment a dataset lands. Figures that would repeat one another are not drawn, and the step says which it skipped and why.
The visual pipeline
330 components: the cleaning, statistics and machine learning every study uses, and nine research domains down to differential expression, gene set enrichment and protein interaction networks. Python or R, on libraries the platform has tested and approved.
Statistics that answer for themselves
t-tests, ANOVA, chi-square, correlation, normality and regression, each drawing its own figure and stating what it found. A test that cannot run on your columns says so plainly instead of returning a number.
Models, and the honesty around them
Classification, regression, clustering, semi-supervised learning and ensembles, with hyperparameter search. Splits detect repeated subjects and group them, because a leaked patient is an accuracy score that means nothing.
AlchemYst-Crucible
Our reasoning agent turns your canvas into runnable code: it plans the analysis, writes the harder parts in parallel, then assembles and reviews the result against your data before any of it runs. You can read, edit and export everything it writes.
Lineage that survives the viva
Every run keeps its code, parameters, environment, data version and outputs, and every agent action lands in an append-only log. Reproducing a result from eighteen months ago is opening it, not rebuilding it.
Reports, models and publication
Drafted reports you edit and export to PDF or LaTeX, trained models served as endpoints with their own model card, and findings linked across datasets and papers in the knowledge graph.
Who it is for
Four people, one bench
A platform a laboratory can actually adopt has to answer to the person doing the work, the person accountable for it, the group around them, and the institution paying for it. These are not the same question.
Researchers and PhD candidates
You have the question and the data. The infrastructure should not be the hard part.
- Build an analysis without writing the plumbing around it
- Keep every run, figure and parameter without remembering to
- Hand a supervisor something they can open, not a folder of notebooks
Supervisors and principal investigators
See what the group is doing without asking each person what they are doing.
- Progress, pipelines and reports for every project on one page
- Send a to-do against a specific pipeline and watch it close
- Review the work as it stands rather than as it was described
Research laboratories and groups
A shared bench, with the access rules a real group needs.
- Fourteen research roles, from department head to visiting researcher
- Read or edit permissions per feature, approved by a supervisor
- Datasets, pipelines and findings shared across partner laboratories
Institutions, IT and compliance
The questions procurement asks, answered before they are asked.
- EU data residency, and no model trained on your data, ever
- Single sign-on, directory provisioning, and access that ends when someone leaves
- An append-only audit trail, a signed processing agreement, and your own instance if you want one
Built for computational science
Why it is built this way
Reproducibility is the default, not the discipline
In Nature’s 2016 survey of 1,576 researchers, more than 70% of respondents reported having failed to reproduce another scientist’s experiments. Keeping the analysis record is one practical response: the platform retains the data reference, code, parameters and results as the work runs. Read the survey.
European by construction
Data, compute and inference stay in the European Union. Your research data is not used to train third-party AI models. An institution can be pointed at its own inference account so that even the prompts stay inside its own tenancy.
Python and R, without the environment
The same canvas emits either language and runs it in an isolated sandbox with no network. Nobody installs anything, nobody debugs somebody else's dependency tree, and a pipeline that ran last year still runs.
Built with research, not at it
The platform comes out of a doctorate on data infrastructure for applied science and eighteen years of putting machine learning into operations where being wrong has consequences. Every component here exists because a study needed it.
The question behind the platform
Why this platform exists
Rebuilding an analysis takes its data, code, parameters and context. Our starting point is to keep those together as the work happens.
In Nature’s 2016 survey of 1,576 researchers, more than 70% of respondents reported having failed to reproduce another scientist’s experiments, and more than half their own. Baker, Nature 533, 452–454 (2016).
These are respondents’ reported experiences, not a measured failure rate for all published research.
Watch the platform walkthrough or listen to the research conversation
A walk through the platform
From loading data and building a pipeline to recorded runs, reports and publications.
A conversation about reproducibility
Why analysis context gets lost, and how tools can help keep it. Audio, 19 minutes.
Research beyond the platform
Recent work from the wider research community, curated through OpenAIRE. These are external publications, separate from studies shared on AlkhemYst-Ai.