‘Anyone can download an open-source LLM. But you can’t simply download more than 12 years of specialised staffing data. That’s what makes our LLM valuable to staffing companies,’ said CEO Jan Govaerts at the launch of the LLM developed specifically for staffing and recruitment by theMatchBox, a Belgian AI specialist and part of the international Pixid Group.
The LLM is a so-called open-source base model, trained on more than 12 years of HR domain data.
The new LLM is a so-called open-source base model trained on more than 12 years of HR domain data, and is hosted on European infrastructure. The launch is a direct response to the limitations of generic AI models, which, in Govaerts’ experience, ‘lack sufficient domain knowledge for recruitment’. It also responds to growing market concerns about reliance on external providers for tokens and data storage outside Europe, he says.
Job vacancies provide insights
According to Govaerts, the added value of this specialised LLM lies in the extensive training data fed into it, including job vacancies, skills, occupations and companies. This knowledge enables the model to share unique HR insights that a general-purpose language model certainly cannot yet offer, he says. ‘The model was trained without using client data, to avoid privacy issues,’ he adds.
‘The model was trained without using client data, to avoid privacy issues.’
In addition to the training data, theMatchBox has developed its own ontology that maps the relationships between occupations, skills and job titles. The ontology checks the substance of the LLM’s output, ‘identifying incorrect connections and systematically reducing the risk of bias and AI hallucinations’, according to Govaerts. The LLM incurs no external token costs and runs entirely on European infrastructure. Sensitive information is therefore not sent to external LLM platforms either, the CEO emphasises.
Fast processing
Govaerts sees further advantages, including fast processing because the system is optimised for specific recruitment tasks, lower energy consumption than general-purpose LLMs, and immediate compliance with the EU AI Act. ‘The technology fully meets the applicable requirements of the EU AI Act. Compliance, transparency, data protection and human oversight are integral parts of the architecture,’ he says.
‘So don’t ask our LLM to plan a road trip across the United States…’
According to Govaerts, theMatchBox has deliberately chosen not to compete with the largest general-purpose language models with its own LLM. ‘So don’t ask our LLM to plan a road trip across the United States. Ask it to understand the skills in a CV, interpret a complex job vacancy or recognise relevant connections between candidates and jobs. That’s what we’ve been working on for 12 years,’ says Govaerts. ‘We don’t need the biggest LLM. We need the smartest LLM for staffing.’
International rollout
The launch of its own LLM marks the next step in the development of theMatchBox’s Staffing Intelligence solution. As part of the international Pixid Group, the company now also has the scale, expertise and international market knowledge to make its AI technology available to staffing organisations worldwide, Govaerts notes. ‘This enables us to deliver specialised, reliable and compliant AI capabilities that help staffing companies work more intelligently, quickly and efficiently.’