Welf Labs / Research agenda

Intelligence that learns from industry.

Welf Labs explores how AI can understand situations, adapt to change and learn from experience. Our research is shaped by the demands of the physical world.

Start a conversation

Humanoid with paired front-facing camera lenses and a central Welf chest emblem in neutral studio light

The perspective

The real world does not stand still.

Industrial systems face changing materials, unfamiliar conditions and incomplete information. We study how intelligence can respond to that variation, then bring the research into practical engineering.

Welf Labs

Four questions for the next intelligence.

Situational awareness

Understand the operating situation.

A machine alarm, an inspection image and yesterday’s work order tell different parts of the story. Our research asks how AI can combine them into an accurate view of what is happening now.

Research approach & applications

The research question

How can a model connect observations to the right asset, follow changes over time and recognise when evidence is stale, incomplete or contradictory? We study multimodal context, task memory and uncertainty.

Industrial applications

Maintenance triage in energy and process plants, aerospace service preparation, maritime machinery inspection and manufacturing changeover checks. The useful output is a supported account of the situation and the information still needed.

Evaluation criteria

Reconstruct operating states from timestamped records and images. Measure missed changes, unsupported conclusions, source accuracy and the time an operator needs to review the result.

Adaptive intelligence

Respond when conditions change.

A new part variant, an unavailable tool or a revised instruction can break a workflow that worked yesterday. Adaptive intelligence concerns how the system responds during the task.

Research approach & applications

Adapt the plan to the situation

Research how agents recognise a change, select useful context and revise the next step. Compare fixed workflows with systems that can replan, use another permitted tool or ask an operator for help.

Transfer across industrial settings

Investigate robotic handling across part variants, manufacturing changeovers, shipping exceptions and maintenance planning. Study which skills transfer between sites and which need additional examples.

Measure recovery, not just completion

Test unfamiliar layouts, missing inputs, tool failures and changed priorities. Compare completion rate, recovery time, repeated actions, operator interventions and the cost of adapting.

Explore robotics and physical AI

Self-improving AI models

Learn from verified experience.

An expert corrects an answer. A completed task reveals a better method. We study how that experience can improve the next model while preserving what it already does well.

Research approach & applications

Learning methods

Compare fine-tuning on reviewed corrections, preference learning, distillation and reinforcement learning where outcomes can be reliably scored. Study how repeated updates affect earlier capabilities.

Industrial applications

Technical-document extraction, maintenance recommendations, recurring SOP exceptions and robot skills learned from demonstrations and operator interventions.

Evidence across model versions

Keep unseen cases and earlier tasks outside the learning loop. Measure improvement, regression and review effort across updates. Compare changes to model weights separately from changes to retrieval, prompts or tools.

Explore SOP automation

Specialised industrial LLMs

Work with the language of an industry.

Part numbers, tolerances, material grades and procedure revisions carry meaning that general answers can miss. Our agenda examines how language models acquire that knowledge and apply it to a defined task.

Research approach & applications

Choose the right form of specialisation

Compare retrieval, fine-tuning and their combination. Investigate smaller models, German and English technical language, and multimodal extensions for drawings and tables. Assess continued pre-training where the data and results justify it.

Engineering and manufacturing

Extract RFQ requirements, reconcile specification changes and interpret material certificates. Test units, tolerances, component identifiers and source revisions across automotive, manufacturing and metals workflows.

Energy and regulated operations

Retrieve asset-specific maintenance procedures in energy, oil and gas, petrochemicals, aerospace and maritime operations. Investigate SOP retrieval and deviation-evidence preparation in pharmaceuticals.

Evaluate the complete task

Compare domain accuracy, source support, reviewer corrections, latency and cost per accepted result. Test on document families the model has not seen, including missing information and conflicting revisions.

Explore private-cloud AI

Two researchers exchange ideas beside a machine-vision camera against a neutral studio background

Engineering note

Evaluation before autonomy.

How to build a representative test set, measure results against the current process and decide when a workflow is ready for more automation.

Read the note

A closer look

What goes into the work.

Explore the questions that matter for your project.

Make progress measurable.

Each research question needs a defined task, a credible baseline and evidence that survives a change in conditions.

A representative test

Use authorised industrial examples with domain review. Separate development data from unseen assets, document families, sites or time periods.

A useful comparison

Compare methods with equivalent information, tools and budgets. Report quality alongside review effort, recovery time and operating cost, including the difficult cases.

A method others can examine

The intended outputs are technical reports, evaluation sets and reference implementations. Record the conditions under which a result holds so a delivery team can assess it for its own workflow.

Bring an industrial research question.

A difficult operating condition, a specialised dataset or a recurring failure can be the starting point. Tell us what the system needs to understand and how you would judge a useful result.

Start a conversation