When Video Becomes Machine Knowledge

How Slick+ proved a multilingual AI pipeline on Europe’s LUMI-G supercomputing infrastructure – and de-risked the next step towards market

Every organisation has a second knowledge base. It is not sitting in a document repository or an LMS. It lives in the heads, habits and improvised explanations of the people who keep operations moving.

At Slick+, we help frontline teams capture that know-how as short video: how a task is really done, what changes on a night shift, which detail prevents a repeat error, or how a site standard should look in practice. Yet a growing library of clips can become another silo unless its contents can be understood, translated, structured and found.

Through FFplus, Slick+ is building an open-source-model route from raw video to multilingual, machine-searchable operational knowledge. The latest project milestone moves that proposition beyond a diagram. Deliverable D1.1 confirms that Task 1 is complete: HPC access has been established, the Slick+ AI pipeline has been connected to LUMI-G in Kajaani, Finland, and the first proof-of-concept and benchmark runs have been executed successfully.

The team secured a 10,000 GPU-hour EuroHPC AI Factory Fast Lane allocation for 23 June to 23 September 2026. It configured the LUMI-G environment, generated language-specific model checkpoints, automated the cloud-to-HPC file-transfer path, put Llama-based feature extraction in place, and began Task 2 ahead of schedule.

For customers, partners and investors, that is the important change in status: the core research workflow now runs.

This is no longer only a pipeline design. It is a working research system with measured training runs, generated checkpoints, automated data movement and structured feature extraction already in place.

What has now been proven

Enterprise video is a demanding AI input. Speakers bring different accents and languages. Operational vocabulary is specialised. Recordings may come from noisy warehouses, factories, construction sites or other frontline environments. A transcript is useful, but it is not yet a reusable knowledge object.

The commercial proposition depends on a chain of capabilities-ingestion, speech recognition, translation, structured metadata and reliable return of results-working together.

Five language models trained in parallel

The Task 1 proof of concept used the open-source Whisper Small model. Small was chosen to validate speed, compatibility and pipeline integrity; it is not the final production model. The team fine-tuned the model on Google’s FLEURS speech dataset for English, Spanish, French, Arabic and Mandarin.

Five Slurm array jobs-one per language-ran simultaneously on LUMI-G and completed in approximately 1,500 seconds, or 25 minutes. Each produced a language-specific checkpoint for the next phase of inference and benchmarking.

This is a throughput and workflow-validation result, not yet an accuracy result, but it demonstrates the repeatable parallel training pattern the project needs.

Inference and single-step translation executed

The English checkpoint was tested against a real-world radio interview with multiple speakers and mixed accents, and 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, without requiring a separate transcription-then-translation stage. That is important for Slick+ because frontline knowledge does not always move in one language at a time.

D1.1 does not yet report a measured word-error rate or full end-to-end production latency. Those remain questions for the formal benchmark phase.

An automated cloud-to-HPC transfer path is working

The transfer path between Slick+’s cloud environment and LUMI-G is already automated. AWS / GC is polled for new files, and transfer of those new files takes place through a scheduled Slurm job. The file is staged in scratch storage, processed through the AI pipeline, and returned for downstream handling.

That matters because the project is not limited to a manual research demonstration. It has already moved towards the operating pattern Slick+ will need in practice: new video or audio inputs can be detected, moved into the HPC environment, processed as batch jobs, and made available for the next layer of enrichment.

Further automation and hardening remain part of Task 2, but the key bridge between the cloud workflow and LUMI-G has been established.

Feature extraction is already in place

The second critical layer is also now in place: feature extraction from transcripts using Llama, served through Ollama. Ollama is the engine used to serve open-source models; Llama provides the model capability. As part of Task 1, this layer converts transcript content into a structured JSON object including title, description, tags, learning goals and required skills.

This is a commercially significant step. Speech recognition turns video into text. Feature extraction turns that text into something an organisation can organise, search, recommend and measure.

For Slick+, it is the difference between a growing video library and a machine-searchable knowledge system.

Under the hood: an AMD-native AI stack

LUMI is an HPE Cray EX system with 380 petaflops of sustained computing power. Its GPU partition, LUMI-G, uses AMD rather than NVIDIA accelerators. Each LUMI-G node combines a 64-core AMD EPYC CPU with four AMD MI250X GPU cards, exposed to Slurm as eight GPU devices.

The Slick+ team validated PyTorch 2.x, ROCm 6.x, Hugging Face Transformers 4.x, Whisper and the Ollama-served Llama feature-extraction layer in the LUMI environment. Using the supported container stack avoided rebuilding the software environment from scratch and reduced setup risk.

The operating model is deliberately different from a public cloud console. While some HPC compute environments restrict public internet access from compute nodes, LUMI compute nodes have public internet access. The architecture still demands discipline: code, models and data must be managed through the correct HPC workflow; durable project storage is separated from high-volume scratch space; critical outputs are backed up externally; and Slurm schedules interactive or batch work onto available compute resources.

Those constraints matter commercially. A service that works only as a manual notebook experiment is not a product. Turning it into reproducible, discrete batch jobs-with automated data movement, explicit staging, checkpoints, monitoring and failure handling-is the engineering bridge between access to powerful GPUs and a dependable customer workflow.

Support from the MINERVA AI-on-HPC service helped the team validate this architecture and prepare for multi-node optimisation in Task 2.

Why LUMI changes the economics of experimentation

LUMI is powered by 100% renewable electricity, with waste heat reused in Kajaani’s district heating network. For an SME, its most immediate value is access: repeated training, inference and evaluation can be explored without turning each experiment into an open-ended commercial cloud bill.

D1.1 estimates that Task 1 used only around 10–20 GPU hours of the 10,000-hour allocation, leaving substantial capacity for the larger Whisper large-v3 runs planned in Task 2 and later experimentation.

The point is not that HPC makes AI free. It gives the team room to test model choices, language coverage, orchestration and deployment patterns before production unit economics are locked in.

The business case: control, reach and speed

For customers, the first benefit is discoverability. A video that can be automatically transcribed, translated, titled, tagged and linked to goals or skills is far easier to reuse than an anonymous file in a folder. That matters when an organisation needs to find a safe method, onboard a new colleague, transfer knowledge before an experienced employee leaves, or compare practice across sites.

The second benefit is reach. A multilingual pipeline can reduce the distance between the person who holds the knowledge and the colleague who needs it. In frontline environments-where workers may not share a first language, a desk or a company email account-that supports operational consistency, inclusion and risk reduction.

The third benefit is strategic control. Slick+ is proving a path based on open-source and open-weight local models rather than depending entirely on proprietary model-as-a-service endpoints. Commercial AI services such as OpenAI or Anthropic can be powerful, but relying on them for every inference call can create constraints around cost, data flow, deployment choice, model behaviour and customer assurance.

By building capability around portable models that can run in controlled environments, Slick+ gains more influence over quality, governance and unit economics.

That creates optionality. The same underlying AI workflow can be developed and benchmarked on LUMI, moved to scalable cloud infrastructure where that is the right commercial model, or deployed closer to the customer through local or on-premise solutions where data control, latency, cost or procurement requirements make that preferable.

For enterprise buyers, that flexibility is part of the product value.

A disciplined route from research to revenue

Progress was not friction-free. The initial Fast Lane application required resubmission, and an unanticipated federation-platform identity step pushed D1.1 approximately three weeks beyond its planned due date. The deliverable reports no impact on the December D2.1 milestone. Task 2 has begun ahead of schedule, and a follow-up allocation application is planned before the current access window expires.

The roadmap is concrete: 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’s target is a 60% improvement in non-English transcription, but that is a forward benchmark-not a result claimed in Task 1. BERT-based sentiment and personalisation work is planned after the transcription workflow is automated.

LUMI remains the environment for research, proof and optimisation, not necessarily the final revenue-generating host for every deployment. The immediate route to market is controlled pilot work with selected clients. If those pilots validate customer value, the optimised models and automated pipeline can move to paid, scalable infrastructure such as AWS or an equivalent commercial environment, or to local/on-premise solutions-including Mac Studio-class deployments-where that better fits customer needs.

This is not supercomputing replacing the cloud or the customer environment. It is each environment doing the job it is best suited to do: LUMI for high-performance experimentation and benchmarking, cloud for scalable managed delivery, and local/on-premise infrastructure for organisations that need tighter control over data, cost, latency or deployment.

Trust is part of the product

Technical progress has been paired with governance. 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. The basis for that assessment is important: Slick+ is a knowledge-sharing and learning tool, 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, that sequence is material: prove the system, define the controls, then introduce sensitive data under an agreed governance framework.

A new interface to frontline intelligence

The most valuable knowledge in an organisation is often the least structured. It is spoken, demonstrated and adapted in context. It lives with the people closest to the work. Until now, that has made it difficult for machines to index and organisations to scale.

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

The next opportunity is bigger than transcription. It is to turn frontline video into structured, multilingual, searchable machine knowledge-and to 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.

 08 Sept 2026
For more information, please get in touch with paul@slick.plus