“Our job is quite simple: to help you find the perfect piece of furniture easily,” she sums up from the outset. The idea for Desai was born of a trivial observation for anyone who has ever tried to fit out a real flat with very real constraints. “Instead of spending a year or two trying to find the perfect image, whether you have a dog or a corridor that’s a little too narrow, we help you find the ideal room. It’s a bit like a sofa Tinder,” she says, smiling.

Behind the formula lies a major technological challenge. Desai first explored existing generative AI models. “We started by testing GPT and other models to see if they understood furniture, but they were very good at inventing it, not understanding it,” she explains. The verdict was clear: only around 50% accuracy. Insufficient for an application that is supposed to recommend a piece of furniture that has to fit through a door, fit into a specific space and respond to very concrete usage constraints.

Coming from the world of computer vision, Crystal Van Oosterom knew she had to change her approach. Desai set about training her own models, using some 300,000 hand-labelled images. The limitations of the conventional infrastructure soon became apparent. “We were crashing AWS machines because our models were too heavy.” The start-up needed raw power, and fast.

This is where Meluxina comes in. “We needed enormous computing power. As a Luxembourg company, being able to do this locally was invaluable, and access to Meluxina via the programme was free.” The leap was immediate: accuracy rose from 50% to over 80%. Today, Desai works with more than 90 million data points. “We need a hell of a machine to drive all this in a cost-effective and efficient way. Meluxina offered this power and the necessary technical expertise.”

Beyond the technological performance, the founder insists on the structuring imposed by the Luxembourg ecosystem. “I’m not known for my patience”, she confides. However, the R&I programme forced the team to slow down and focus on its needs. “It forced us to take a step back to structure our intellectual property and our technical needs, for example how to measure a part to an accuracy of two centimetres. This work was initially carried out without a dedicated engineer.

The institutional support leaves a lasting impression on her. “This is the first government I’ve seen send emails on a Sunday afternoon or Monday at 7am to point out a typo in an Excel spreadsheet.” Drawing on her experience in venture capital, she decides: “To have a supercomputer in your back garden and such government support is an incredible opportunity that you can’t find in London or Switzerland.”

When asked by the public about model version management, the answer is very operational. “Meluxina offers backups, but we also download each version, even each training era, locally.” An essential precaution: “Sometimes the model degrades after a while, so you need to be able to go back and find the sweet spot.”

On the rampant evolution of open source and AI models, Crystal Van Oosterom is surprisingly pragmatic. “I read a comparison with iPhones: at the beginning, the changes between each model were massive, then the evolution slowed down. It’s the same for AI models.” He concludes, addressing the entrepreneurs present: “If your current model works perfectly and solves your problem, you don’t need to chase the next shiny thing just for the sake of it.”

From four-dimensional data to forest well-being

In an evening dedicated to the computing power and promise of HPC, Desai’s testimony served above all as a reminder of one thing: behind supercomputers and millions of data points, innovation sometimes starts with a very simple problem - finally finding the right sofa. Beyond the case of Desai, the discussions quickly showed that the use of intensive computing is neither reserved for generative AI nor confined to a single sector. Around the table, three other entrepreneurs illustrated the diversity of uses for Meluxina and, above all, the same common thread: the need to move fast, without needlessly reinventing the wheel.

At Space Time, Shahriar Agaajani recounts a trajectory made up of successive adaptations. The start-up works on so-called "four-dimensional" data, combining space and time, to extract usable information. Initially focused on construction sites, the company has had to review its priorities as crises have unfolded. “We were hit by crisis after crisis, by Covid and the rising cost of materials, all of which delayed projects,” he explains. The turning point came almost by chance, while observing LiDAR images of forests. “I realised how rich the data was: number of trees, biomass, etc. We turned to forestry, then mining and safety. An ability to adapt that he sums up with a quote often attributed to Darwin: it’s not the strongest who survive, but the one who adapts best.

On the technological front, pragmatism takes precedence. “We had our own algorithms, but LuxProvide’s computers were much more powerful,” he says. Rather than rebuild everything, the team chose to rely on existing open source bricks. “They were already doing 60-70% of the work. All we had to do was plug in and play. It was very cost-effective and we saved a huge amount of time.”

Similar logic at Symbiose Management, where Marwen Nefati aims to create a “Bloomberg of the forest”. The company combines satellite data, LiDAR and other sources to analyse forest ecosystems, assess their economic value and anticipate their evolution in the face of climate change. This positioning places the company at the heart of natural capital issues, but also means it has to handle massive volumes of data. “We’re not talking Excel here, but gigabytes, even terabytes”, he points out. “I’ve seen my memory requirements explode to over 90 gigabytes, causing my previous systems to crash. If the infrastructure isn’t designed to be scalable from the outset, you’re stuck technically and financially.”

Meluxina, a talent magnet

Access to a supercomputer then becomes a strategic argument, including for recruiting. “Having the word ‘HPC’ in a job advert changes everything,” observes Marwen Nefati. Profiles from large laboratories, sometimes reluctant to join a private start-up for fear of losing access to this type of tool, are suddenly showing an interest. “Someone from NASA joined us for just this reason.”

In a different vein, Loveneesh Rana, founder of Spider Space, is working on automating operations in the space industry using artificial intelligence agents. For him, the difficulty lies not only in the performance of the models. “It’s a trick question to ask whether the hardest part is building the model or integrating it,” he smiles. “Both are difficult.” Beyond the AI itself, the architecture, the coordination between several agents, the calling of tools and the user interface are decisive. “The value is as much in the model as in the final solution that actually solves the problem.”

After the experimentation phase, Spider Space is now exploring two avenues: training a multimodal foundation model dedicated to space, developed in Luxembourg for the European ecosystem, and marketing its current agentic solutions, keeping an eye on the future evolution of the infrastructure, in particular Meluxina 2.0.

Through these very different backgrounds, the same message emerges from the evening: supercomputing is not an end in itself, but an accelerator. Whether it’s measuring a part to within two centimetres, mapping a forest or automating a space factory, Meluxina appears to be a tool for scaling up, provided it is used at the right time and with sufficient technical maturity. A logic summed up by the LuxProvide host in conclusion: there is no magic moment to switch to HPC, but a need to be ready to test, to fail sometimes, and to learn quickly.