Research and development

Our agents locate the elements on a screen to click and type in place of a person, including in software with no API.

We adapt large open models to forecasting, energy production in particular, and test them on data they have never seen.

Our models treat each page as an image, diagrams, tables, scans and manuscripts included, and find the right page among thousands.

We study how AI engines choose the sources they cite, and have published a content score whose resistance to manipulation we measure.

We work on how teams of agents cooperate in production, and on checking what an agent claims to have done.

We study where agents that write reports, spreadsheets and presentations go wrong, and how to measure the quality of what they deliver.

We train small guardrail models that an organisation can host and control itself.

We specialise open models so that a small model is enough for a specific task: query routing, reinforcement learning, domain adaptation.

We measure what an assistant grounded in the institution's own courses adds over a general-purpose model, and design tutors that guide students without giving them the answer.

We rebuild in the browser the layout of a word processor and the functions of a spreadsheet, and measure how close we come to Word and Excel.