Innovation within Bounds
summary
Modernization targets a 1970s-born estate that still pays ~640 billion kronor yearly, so migration choices are pragmatic: legacy COBOL/C++/Java components move to OpenShift and containers while some systems remain frozen. The team data-mines the entire codebase into a graph to map dependencies, runs a Swedish language model fine-tuned in PyTorch on GPUs for medical-text classification (first production model 2019), operates a production ML pipeline with provenance for the EU AI Act and GDPR, and ships a pii-mos PII mask tool. The claim: training data and traceability, not models, are the hard part.
AI-written summary · about
chapters (14)
- 0:00Innovation within constraints
- 2:12Technology evolution and legacy systems
- 6:50Retiring outdated systems
- 7:58Legislation and policy changes
- 10:54Organizational inertia and Diamond model
- 14:27Data migration challenges
- 19:15Flexibility versus structure in data
- 21:50Knowledge base for data management
- 22:35Graph technology for code complexity
- 24:02AI adoption and investments
- 25:48AI for disability assessment
- 31:40ML pipeline and AI Act compliance
- 33:08Other AI applications and tools
- 36:44Q&A on Linux versions
related talks
55mThe healthcare evolution: Modernize systems for operational excellence
34mScaling ML in the Traditional Enterprise
28mNavigating LLM deployment: Tips, tricks, and techniques
29mMachine learning in production
20mAccelerating innovation using tiered AI Architecture
33mDealing with the versioning of production-ready models
44mBuilding Multi-Tenant Compute Systems in the Enterprise
52mA MLOps Platform for All Data Personas