How Slick+ used Europe’s LUMI-G supercomputer to test a new way of turning frontline video into searchable, multilingual workplace knowledge

Every organisation has a second knowledge base.

It is not stored neatly in a document folder or learning system. It sits with the people who know how work really gets done: the supervisor who explains the same task ten times a week, the experienced operator who knows which small detail prevents a repeat error, the site lead who can show what “good” looks like better than any policy document can describe it.

Slick+ helps organisations capture that knowledge as short videos. A colleague records how to complete a task, explain a process, share a site update, or show a standard in practice. The result is a video library built from real operational knowledge.

That solves one problem, but it creates another.

As the library grows, people still need to find the right guidance quickly. They may need it in another language. Managers may need to organise it by task, site, skill, risk or learning goal. New starters may need the right video before they begin work, rather than after a mistake has already happened.

This is where the FFplus project comes in.

Through FFplus, Slick+ is testing an open-source and open-weight AI route from raw video to searchable, multilingual workplace knowledge. The latest milestone confirms that the core pipeline is now running on LUMI-G, Europe’s supercomputing infrastructure in Kajaani, Finland.

Task 1 is complete. The team has secured access to 10,000 GPU hours, connected the Slick+ AI workflow to LUMI-G, trained language-specific speech models, automated the cloud-to-HPC file-transfer path, and put Llama-based feature extraction in place.

For a commercial audience, the point is simple: this is no longer a diagram of how the system might work. It is now a working research system that can take video or audio content, process it through a high-performance AI pipeline, and return structured outputs that can be used inside the Slick+ product.

The business problem: video alone is not enough

Frontline organisations often depend on information that is hard to write down.

A warehouse process may change slightly between shifts. A construction site may have local rules that do not appear in a generic induction. A maintenance task may include small judgement calls that experienced people make automatically. A multilingual team may need the same safety or process guidance in a form that does not rely on long written documents.

Video is a strong format for this kind of knowledge because it shows the work in context. People can see the tool, the site, the movement, the standard and the mistake to avoid.

Yet video becomes much more valuable when the system can understand what is inside it.

A useful AI pipeline needs to answer practical questions:

• What is this video about?
• Which language is being spoken?
• Can the transcript be translated?
• Which task, skill or learning goal does it relate to?
• Can the right person find it later?
• Can it be linked to onboarding, safety, compliance or operational improvement?

The FFplus work is focused on that chain, from video and audio input through to structured, searchable knowledge.

What Slick+ has now proven

The Task 1 proof of concept used Whisper Small, an open-source speech recognition model. It was chosen because it is suitable for testing speed, compatibility and workflow integrity. It is not the final model intended for production-quality benchmarking.

The team fine-tuned the model on Google’s FLEURS speech dataset for five languages: English, Spanish, French, Arabic and Mandarin.

Five Slurm array jobs, one per language, ran at the same time on LUMI-G. They completed in about 1,500 seconds, or 25 minutes. Each job produced a language-specific model checkpoint that can be used in the next phase of inference and benchmarking.

This does not yet prove final transcription accuracy. It proves something earlier in the chain: the pipeline can train multiple language models in parallel, produce usable checkpoints, and repeat the pattern at speed.

That is the foundation needed before Slick+ moves to larger models and more formal accuracy testing.

From speech to translation

The team also tested inference, which means using the trained model to process new audio.

An English checkpoint was tested against a real-world radio interview with multiple speakers and mixed accents. It returned a full transcript in the test run. The recorded delay from Slurm submission to GPU allocation was about 15 seconds.

A separate test processed Spanish speech directly into English text in one Slurm job. That avoids a separate transcription step followed by a separate translation step.

For Slick+, this is a practical capability. Frontline knowledge often sits across languages. A colleague may explain a process in Spanish, Polish, Arabic or Mandarin, while another colleague or manager needs to find and use that knowledge in English or another working language.

The formal benchmark phase still has work to do. D1.1 does not yet report measured word-error rates or full end-to-end production latency. Those will be tested as the project moves from the proof-of-concept model to larger multilingual benchmarking.

The file-transfer path is already automated

A key part of this work is not the AI model itself. It is the movement of files through the system.

In a live product, customers will not upload files manually to a research environment. The system needs to detect new content, move it to the right processing environment, run the AI job, and return the result.

That path is now in place.

Slick+’s cloud environment is polled for new files. When new content is detected, the transfer takes place through a scheduled Slurm job. The file is staged in scratch storage, processed through the AI pipeline, and returned for downstream handling.

This means the project has already moved beyond a manual lab test. New video or audio inputs can be detected, transferred, processed as batch jobs, and prepared for the next stage of enrichment.

Task 2 will continue to harden this workflow, but the main bridge between the cloud workflow and LUMI-G has been established.

Feature extraction is also in place

Transcription is only one part of the value.

A transcript tells you what was said. It does not automatically tell you how to organise that knowledge, which skill it relates to, or where it belongs in a learning or operational system.

As part of Task 1, Slick+ has put feature extraction in place using Llama, served through Ollama. Ollama is the engine used to serve open-source models. Llama provides the model capability.

This layer converts transcript content into structured fields such as title, description, tags, learning goals and required skills.

For a customer, this is the step that makes the video library easier to use. A short video can move from being a standalone clip to being part of a searchable knowledge system. It can be organised, recommended, measured and linked to real workplace needs.

That is the commercial value of the pipeline: it helps turn video into structured operational knowledge.

Why LUMI-G is useful for an SME

LUMI is an HPE Cray EX supercomputer with a large GPU partition. It is powered by 100% renewable electricity, with waste heat reused in Kajaani’s district heating network.

For Slick+, the commercial value is access to high-performance experimentation without turning each test into an open-ended cloud bill.

AI development often requires repeated testing. Teams need to compare model sizes, languages, datasets, training runs, inference performance and deployment choices. In a commercial cloud environment, those experiments can become expensive before the business case is fully proven.

LUMI-G gives Slick+ room to test before locking in production choices.

Task 1 used only around 10 to 20 GPU hours from the 10,000-hour allocation, leaving significant capacity for larger Whisper large-v3 runs and later experimentation. The aim is not to make AI free. The aim is to reduce the risk of choosing the wrong model, workflow or deployment route too early.

Open models give Slick+ more control

One of the most important commercial questions in AI is where the model runs.

Many companies use proprietary model-as-a-service tools, such as OpenAI or Anthropic. These services can be powerful and useful. They can also create constraints around cost, data flow, deployment options, customer assurance and long-term unit economics.

Slick+ is testing a different route: open-source and open-weight models that can run in controlled environments.

That gives the company more options.

The same workflow can be developed and benchmarked on LUMI-G, moved to a scalable cloud environment when that is the right delivery model, or deployed closer to the customer through local or on-premise infrastructure where that is required.

For some customers, a cloud-based model may be fine. For others, data control, procurement rules, latency, cost or security expectations may point towards a local deployment. Mac Studio-class infrastructure is one example of how smaller local deployments may become part of the practical route to market.

This flexibility is commercially useful. It means Slick+ is not tied to one AI vendor, one hosting model, or one cost structure.

What this means for customers

For customers, the first benefit is discoverability.

A video that can be automatically transcribed, translated, titled, tagged and linked to goals or skills is easier to reuse than an anonymous file in a folder. That helps when an organisation needs to onboard a new colleague, explain a safe method, transfer knowledge before someone leaves, or compare how work is done across sites.

The second benefit is reach.

A multilingual pipeline can reduce the gap between the person who holds the knowledge and the colleague who needs it. In frontline environments, workers may not share a first language, a desk, or a company email account. Short video, supported by transcription and translation, can make practical knowledge easier to access.

The third benefit is control.

Customers increasingly want to know how AI is being used, where their data goes, what models are involved, and whether the system can meet their governance needs. A pipeline built around portable models gives Slick+ more room to meet those requirements.

The route from research to revenue

The project has had some friction. The initial Fast Lane application required resubmission. An unanticipated federation-platform identity step pushed D1.1 about three weeks beyond its planned date.

The deliverable reports no impact on the December D2.1 milestone. Task 2 has already begun, and a follow-up allocation application is planned before the current access window expires.

The next work is clear: continue hardening the automated cloud-to-LUMI file exchange, move from Whisper Small to Whisper large-v3 for higher-quality multilingual benchmarking, extend and validate the Llama-based feature-extraction service, and use MLflow to compare accuracy across training iterations.

The project target is a 60% improvement in non-English transcription. That is a future benchmark, rather than a result claimed in Task 1. BERT-based sentiment and personalisation work is planned after the transcription workflow is automated.

LUMI-G remains the research, proof and optimisation environment. It is not necessarily where every customer deployment will run. Once pilots validate customer value, the optimised models and workflow can move to the right delivery environment: cloud, local infrastructure, on-premise hardware, or a mix of these.

Trust and governance

The technical work has been matched with governance work.

A review led by CSC’s Senior Coordinator for Trustworthy AI concluded that the described Slick+ use case does not currently appear to be high-risk under the EU AI Act.

That assessment is based on the role of the product. Slick+ is designed as a knowledge-sharing and learning tool. It is not a system for recruitment, employee evaluation, performance monitoring or discipline.

Compliance mode is transparent and consent-based. Standard mode personalisation does not generate individual performance scores or management-visible engagement data. Employees can opt out of personalisation and browse content freely.

Task 1 and Task 2 training use public FLEURS data and dummy or synthetic inputs. A Data Processing Agreement with the Task 3 enterprise pilot partner will be executed before any real employee data is processed on LUMI-G infrastructure.

For enterprise buyers, this sequence is important: prove the system, define the controls, then introduce sensitive data under an agreed framework.

The opportunity

The most valuable knowledge in an organisation is often the least structured. It is spoken, shown and adapted in context. It lives with the people closest to the work.

Until now, that has made it hard for software to index and hard for organisations to scale.

Slick+ has now shown that its multilingual pipeline can run on European supercomputing infrastructure, with automated data movement and Llama-based feature extraction already in place as part of Task 1.

The opportunity is larger than transcription. It is to help organisations turn frontline video into structured, multilingual, searchable knowledge, then deliver it through the right mix of supercomputing research, scalable cloud delivery and controlled local deployment.

Funding acknowledgement: This project has received funding from the European High-Performance Computing Joint Undertaking under grant agreement No 101163317. The Joint Undertaking receives support from the Digital Europe Programme. 

For a more technical deep dive, please click here.
For more information, please get in touch with paul@slick.plus