AI work done by the people who do the research.
We are researchers in mathematics, physics, economics, statistics, finance and machine learning. We teach organisations to use AI on their own data, and we build the expert data that AI models are evaluated and trained on. We take on few clients at a time, so the people who design the work are the people who do it.
AI capability programmes for institutions and companies
Forecasting and data analysis, taught on your own records in three residencies, until the methods are yours.
How a programme runs Evaluation & post-training dataExpert data for AI models
Evaluation sets, red-team sets and post-training data, written and graded by academic specialists.
What we buildAI in practice, taught by the people who do the research. Built around your work.
We design and deliver AI programmes in forecasting and data analysis around your own records, workflows and questions, and we stay with you through three rounds of application, until the methods are yours and not ours.
The designers are the teachers
The researchers who design your programme are the ones in the room, every session.
No programme is taught twice
Every exercise is written on your data and your cases. We take on few clients at a time so that this remains true.
Judgement, not just tools
We teach when a forecast can be trusted, what an uncertainty band means, and where an AI answer must be verified before it is used.
How a programme runs
Three residencies on your premises, separated by periods in which your team applies the methods in their day-to-day work and logs what comes up.
Foundations, on your data
Methods, tools and safe use, taught on your own series and records from the first morning.
Everything that came up
The programme is rebuilt around the questions, blockages and new use cases from the application period.
Consolidation and hand-over
Open questions closed, procedures written down, models and assistants documented.
Groups of up to 20 · technical or management track · delivered in English, with materials in Arabic or other languages on request · theory 20–30%, practice 70–80% · residency length and spacing set with you during design and assessment.
The programmes
Two tracks. Each programme can be delivered as a technical course for the people who will build and run the models, or as a management course for those who commission the work and act on the results. Same cases, different depth; a client can run both.
Forecasting with AI
For planning, finance, operations, policy and strategy units: anyone who has to put a number on next quarter and defend it.
- From classical time series to machine-learning and foundation models for forecasting, and when the simple model wins
- Seasonality, trend, structural breaks and regime changes
- Uncertainty: intervals, scenarios, and how to report them to decision-makers
- Evaluation: measuring forecast error properly, backtesting, avoiding leakage
- Language models as assistants for data preparation, code and documentation
- When not to trust a model
- Other topics as your work requires. The final list is set with you during design and assessment.
Your series from day one
The first residency opens with your own data, not a textbook example. By the end of it, each participant has a forecast of something they are responsible for.
Big data with AI
For analytics, data, research and IT units: anyone sitting on large records, whether produced in-house or purchased.
- From raw records to usable data: cleaning, linkage and deduplication at scale
- Extraction from documents (PDFs, scans, free text) with quality control built in
- Analysis with and without code: what AI does for a non-programmer, and where code is still needed
- Statistics that survive scrutiny: sampling, bias, what “significant” actually means
- Recurring reporting and dashboards that run without manual re-keying
- Other topics as your work requires. The final list is set with you during design and assessment.
Your records, in your environment
Exercises run on your infrastructure, on anonymised or approved data, with the tools you are permitted to use.
How we teach
Show
The instructor carries out the task live, on a real document or series from your organisation.
Do it together
The room repeats it step by step on their own machines, with corrections on the spot.
Apply it
Every module ends with a case of the participant’s own.
Consulting: when the work needs doing, not only teaching
Project-based, agreed and priced separately, delivered by the same researchers.
Econometric and economic modelling
Impact and policy evaluation, demand and pricing models, risk and stress scenarios, with methods that stand up to external review and a written account of every assumption.
Forecasting systems
Built on your data, tested, documented and handed over ready to run.
Analysis reports
You bring the question and the data; we deliver a written analysis with the method, the results, and what they can and cannot support.
Your data stays in your environment
We work on approved tools and on anonymised or client-provided data.
Confidentiality
Non-disclosure agreements for everyone involved; material shared to tailor a programme is covered by mutual confidentiality.
Design and assessment first
We meet the unit concerned before writing a proposal; the proposal then fixes modules, timetable and commercial terms.
Depth over volume
Senior people, small groups, long follow-through. We price accordingly, and we do not discount by cutting any of the three.
Next steps
Design and assessment
With the head of the unit concerned: the unit’s work and priorities, the data it may use, the tools it has, and what it needs the programme to deliver. A short questionnaire and one working session follow, from which the programme is written.
A proposal
Modules, timetable, deliverables and commercial terms. A framework agreement where several units are involved.
The first residency
Typically within eight weeks of award, on your premises or at a venue of your choice.
Data for AI models, made by the people who do the research. Built with your team.
For organisations that use or develop AI models, we produce the expert data those models are evaluated and trained on: tasks written, graded and documented by academic specialists in the fields and languages you need. We work alongside your own team from kick-off to hand-over, so that every batch reflects what you need.
Academics, recruited by academics
Our experts are researchers active in their fields, chosen by people who know what a credential means and how to check it.
Few clients at a time
The researchers who design your tasks are on the project from kick-off to hand-over. We take on few clients at a time so that this remains true.
The rigour of a research study
Every dataset is built like a study: a plan for what goes in, written grading rules, two reviews of every task, two quality grades, and a final score. Only tasks above 80% are handed over.
What a dataset can look like
Literature search
The model is asked questions that can only be answered by finding and citing real publications. Experts then check every reference it gives: does the source exist, and does it say what the model claims?
Reasoning and logic
Multi-step problems with worked solutions, graded intermediate steps and notes on where reasoning breaks.
Advanced translation
Specialist texts across languages and registers, graded for fidelity, terminology and tone by native speakers who are also subject experts.
Research questions
Long-form tasks (literature synthesis, method critique, study design) with rubric-graded responses.
Image, audio and document tasks
Annotation, captioning and expert grading of generated output, for models that work on more than text.
Red-team transcripts
Prompts designed to make the model fail on its own terms, the failures it produced, and a taxonomy of why.
What we build
Who this is for. Organisations that use or develop AI models, whether a national or sector model, an in-house fine-tune or a domain assistant, and need evidence of what the model can do, or data to make it better. Language, image, speech or multimodal.
Evaluation sets
To measure what the model can do.
- Domain questions with reference answers and grading rubrics
- Reasoning, logic and quantitative tasks with worked solutions
- Literature search and citation accuracy
- Multilingual and translation quality, graded by native experts
- Long-form research tasks with rubric-based scoring
Delivered with scoring guidance and inter-rater agreement, so a score means the same thing next quarter.
Red-team sets
To find out where it fails.
- Experts probe the model on work they know well and record what breaks
- Factual errors, invented references, reasoning collapses, inconsistency across languages, instruction-following lapses
- Every failure classified in a taxonomy your team can act on
- Severity and reproducibility noted per item
The range is yours to set: from everyday professional tasks to targeted adversarial probing.
Post-training data
To make it better.
- Expert demonstrations of the task done well
- Preference pairs and rankings between model outputs
- Rubric-graded responses with written justification
- Corrections of model output, with the error named
- In the format your training pipeline needs
Built on the same specification as the evaluation set, so improvement can be measured against it.
We specialise in evaluation and post-training data, and take on related projects. For instance, running your model and the leading alternatives through the same expert-graded tasks, to show where it stands against the competition.
How a project runs
One joint team. Your team sees nothing we have not reviewed twice, and nothing goes to volume until your team has reviewed it too.
Kick-off
Your team and ours agree domains, languages, task types, format, volume and timeline.
First batch
Our experts produce a small first batch.
Double review, on our side
Every task is checked by two reviewers, then scored for quality by two graders. Tasks scoring below 80% do not go on to your team.
Your review
Your team reviews the batch against what you actually need.
Adapt and repeat
Specification, rubrics and expert mix are adjusted from your feedback. The same cycle runs for every batch at volume.
Hand over
Datasets, documentation and provenance for every item, with a final round of changes if your review raises them.
What quality means here
- Traceable. Every item carries the expert who wrote it and the review decisions taken on it.
- Measured. Agreement between raters is reported, not assumed; audit samples are drawn at random.
- Reviewable. The documentation is written to pass your internal review, not ours.
Expect changes. We do.
Once real items are on the table, specifications move: a rubric turns out ambiguous, a task type matters more than planned, a model update changes what fails. Re-specification, re-grading and re-evaluation are built into the plan and the timetable, because that is how this work goes.
Commercial logic
Kick-off
Domains, languages, task types, format, volume and timeline agreed with your team.
A specification. No volume yet.
First batch
Small and time-boxed. Double-reviewed by us, then reviewed by your team.
A minimum engagement, per expert-hour. You see real items.
Framework agreement
Volume, pace and terms set on the specification both sides have tested.
Volume, at a pace you set.
Production in batches
Double review, your review, adaptation, delivery, batch by batch until hand-over.
Reviewing each batch; everything else is ours.
Priced per expert-hour, with a minimum engagement. Kick-off and a first batch come before any volume commitment, so that the commitment rests on a specification both sides have tested. The result is a dataset that holds up: every item traceable, every judgement reviewed twice, documentation that still makes sense to whoever audits it in two years.
Everyone signs
Non-disclosure agreements for our team and for every expert, before any access.
Access on your terms
Your model and data are reached only through the arrangement you set. Nothing is copied out; nothing is reused for another client.
Vetted experts
Credentials verified, conflicts of interest checked, performance monitored throughout.
Documentation that passes review
Specification, rubrics, agreement statistics and failure taxonomy delivered with the data, with provenance for every item.
Researchers first
[Field, research focus, and what this person does in a programme or project. Two or three lines.]
[Field, research focus, and what this person does in a programme or project. Two or three lines.]
Our network
We work with a network of PhDs and professors active in research across Europe, in mathematics, physics, economics, statistics, finance and machine learning.
Reach out to us
For a programme, we start with a design and assessment session with the head of the unit concerned. For a dataset, we start with a kick-off meeting with the team that holds the model, followed by a small, time-boxed first batch before any volume commitment.
Delivery in English; materials in Arabic and other languages on request, with native speakers in the network.
Programmes are delivered on your premises or at a venue of your choice.