AI research on the numbers that run your organisation
Kunkafa is an AI research lab focused on tabular and time-series data: the rows and columns in which transactions, operations and measurements are recorded. Kunkafa builds custom models, runs projects on your own data, and advises teams that want to put AI to work on the numbers rather than the words.
Most of the investment went into words. Most of the data is numbers.
Language models have transformed how organisations read, write and search. They were trained on text, and text is what they are good at. The data most organisations actually run on is different: sales by week, transactions by the second, inventory by location, sensor readings by the minute, prices by the tick. It is structured, noisy and numerical, and the same investment has not reached it.
Kunkafa’s research is concentrated on that data. The models are built for structured records and sequences over time, they say how likely an outcome is rather than which outcome sounds right, and they are tested on data they have never seen before anyone relies on them. Language models and numerical models can serve complementary roles inside one organisation. The point is to match the model to the question.
A public showcase you can check
Kunkafa’s market-analysis product is the lab’s research applied to the hardest public setting we could find. Financial markets shift, the data are noisy, and the outcomes unfold in public where anyone can compare a forecast with what happened. Seventy AI experts, trained on about 10 billion data points of price history and tested on about 2 billion they had never seen, each assess how far a price could move, up or down, over any duration from five minutes to seven years. Their agreement and disagreement become Kunkafa’s confidence in each move.
Every forecast is published with both directions and the record beside it. The record is updated each morning and reports whether the level was reached within the stated duration, alongside how many forecasts were counted. Kunkafa is usually right, and the record shows exactly how often. Kunkafa is sometimes wrong, like anyone who forecasts, and the record shows that too.
For an organisation weighing a tabular or time-series project, that record is evidence of how Kunkafa models, evaluates and reports, before any conversation starts.
See the public record · Read how the models work
Services
Every engagement is shaped around the organisation’s own question, data and constraints. The four services below are the forms that work usually takes.
Custom AI models
Models designed, trained and evaluated for one question that matters to your organisation, on your own data.
Demand and sales forecasting, fraud and anomaly detection, risk and pricing models, maintenance and sensor data, and other tabular and time-series problems.
Projects using your own data
A scoped project from data assessment to a model running in your environment, with the evaluation written down before the work starts.
Your data stays yours. Access, hosting and handover are agreed with you in advance.
Consultation for your team
Working sessions with your analysts, engineers and leadership on where AI could help, what the data supports, and what it does not.
Useful before a project, and useful on its own when the question is whether a project is worth starting.
Higher API and MCP limits
Kunkafa’s market analysis inside your own systems and AI assistants, with capacity agreed for your usage.
Every forecast carries both directions and Kunkafa’s confidence, exactly as it appears in the app.
How an engagement runs
- Scoping. One conversation to name the question, who acts on the answer, and what a good result looks like.
- Data assessment. A review of the data you hold: coverage, quality, gaps, and whether it can support the question at all. If it cannot, we say so.
- Success criteria. The measure of success is written down before any model is trained, so the result can be judged rather than argued.
- Modelling and evaluation. Models are built on your data and tested on data they have never seen, in the conditions they are meant to operate in.
- Handover. A model your team can run, monitor and question, with its record kept from the first day.
Why the record matters to a buyer
Moving AI from an experiment into day-to-day operations needs evidence that a model performs in the setting where it is meant to operate, not in a demo. Kunkafa builds every model the way the public product is built: both sides of the question, a measure of confidence, a test on unseen data, and a record kept from the first day. A model that is never scored in public should not be relied upon in private.
Who you would work with
Kunkafa was founded by Kareem Farid, who built Kundera (2019–2022), an early large-language-model application for research and writing, before turning to structured data, and Saleem Mugal, whose long career in consultancy shaped how Kunkafa scopes, evaluates and hands over work. Engagements are led by the founders.
Talk to Kunkafa
Tell us the question, the data you hold and where the answer would be used. Access, capacity and pricing are agreed with you.
Or write to contact@kunkafa.com. To see the models at work first, create a free account.