11/16/2026 updated
MF
not available
info: not available until 12/01/2026
Principal AI Architect / Payments & Banking Architect / Interim CTO
Lidingö (Stockholm), Sweden
Only remote
BSc Business Administration, Linköping University, 2011About me
Principal AI Architect and payments/banking architect with 22 years in cards and payments (Nordea, Swedbank, P27, Zettle by PayPal). Interim CTO and technical leadership, production AI/RAG on self-hosted platforms. Remote across Europe with occasional travel.
Solution Architecture
AI Architecture and Agent Platforms
Design and operation of hybrid AI platforms on self-hosted Kubernetes clusters, combining self-hosted inference and hosted models. Building agent-based systems with MCP tool access, agent identities, permission models and human approval points.
Security, Data Protection and Compliance
About 20 years of PCI DSS experience across card and payment security, combined with GDPR, DORA and DPIA work at major financial institutions including Nordea, Swedbank and P27. Implementation of privacy-first architectures with collector-side pseudonymisation and redaction.
RAG, Retrieval and Grounding
Building retrieval-augmented generation systems with full-text and vector search, validators for cited sources and grounding steps that flag unsupported claims. Use of PostgreSQL, Weaviate and Qdrant as retrieval backends.
Model and Hosting Strategy
Selection of models and hosting per use case based on data sensitivity, cost and required quality. Tracking per-model cost, latency, throughput and error rate against external API pricing to decide between self-hosting and buying.
Evaluation in CI/CD
Controlled agent development where local models review pull requests and larger models review risky changes. Auto-merge only when approved without high or medium findings. Model outputs evaluated against defined RCA cases, scoring overclaiming and restraint when evidence is missing.
Identity and Access Control for AI
Own identities for agents and workloads using OpenBao and OIDC, owner-controlled access, MFA-gated approval flows for selected rotations, and audit without secret values.
Platform and Infrastructure
Kubernetes, Argo CD, GitOps, Prometheus, Loki, Grafana, Python, FastAPI and TypeScript. vLLM and Ollama for local inference, KB-Whisper for speech, and models including Claude, GLM, Grok, Kimi and Qwen.
Solution and Data Architecture
Ownership of target architecture for card data in a domain-driven Data Mesh at Nordea, aligned with emerging AI infrastructure. Writing shared solution concepts for PCI DSS 4.0 and building in requirements from GDPR, DORA, BCBS 239 and AML.
Advising Management and Teams
Translating business goals into architecture decisions for CIOs, boards and engineering teams. Leading AI-assisted development, owning architecture, code review, verification gates and operations.
Design and operation of hybrid AI platforms on self-hosted Kubernetes clusters, combining self-hosted inference and hosted models. Building agent-based systems with MCP tool access, agent identities, permission models and human approval points.
Security, Data Protection and Compliance
About 20 years of PCI DSS experience across card and payment security, combined with GDPR, DORA and DPIA work at major financial institutions including Nordea, Swedbank and P27. Implementation of privacy-first architectures with collector-side pseudonymisation and redaction.
RAG, Retrieval and Grounding
Building retrieval-augmented generation systems with full-text and vector search, validators for cited sources and grounding steps that flag unsupported claims. Use of PostgreSQL, Weaviate and Qdrant as retrieval backends.
Model and Hosting Strategy
Selection of models and hosting per use case based on data sensitivity, cost and required quality. Tracking per-model cost, latency, throughput and error rate against external API pricing to decide between self-hosting and buying.
Evaluation in CI/CD
Controlled agent development where local models review pull requests and larger models review risky changes. Auto-merge only when approved without high or medium findings. Model outputs evaluated against defined RCA cases, scoring overclaiming and restraint when evidence is missing.
Identity and Access Control for AI
Own identities for agents and workloads using OpenBao and OIDC, owner-controlled access, MFA-gated approval flows for selected rotations, and audit without secret values.
Platform and Infrastructure
Kubernetes, Argo CD, GitOps, Prometheus, Loki, Grafana, Python, FastAPI and TypeScript. vLLM and Ollama for local inference, KB-Whisper for speech, and models including Claude, GLM, Grok, Kimi and Qwen.
Solution and Data Architecture
Ownership of target architecture for card data in a domain-driven Data Mesh at Nordea, aligned with emerging AI infrastructure. Writing shared solution concepts for PCI DSS 4.0 and building in requirements from GDPR, DORA, BCBS 239 and AML.
Advising Management and Teams
Translating business goals into architecture decisions for CIOs, boards and engineering teams. Leading AI-assisted development, owning architecture, code review, verification gates and operations.
Experience
Professional experience10+ years
Freelancer since2006
Languages
EnglishFluentSwedishNative speaker
Project history
Root cause analysis for observability data. Designed for customers with strict confidentiality requirements. A separate collector VM pseudonymises entity names and removes free text before data crosses to the analyzer VM, where local models run on own GPUs. Deterministic pseudonymisation plus a local scrubber model on the collector side. Staged analysis from evidence candidates and observations to theories, hard findings and quality reviews, on a source-agnostic model in PostgreSQL with FastAPI and React.
Agent-based operations and root cause analysis. Alerts, deployments, security findings and traces feed an RCA pipeline with an agent loop and its own MCP server. Designed and implemented the permission model: analysis and execution as separate agents, Kubernetes access through a read-only proxy, and risk levels in code that leave restarts, scaling, deletions and pushes to main to approval. An independent model reviews diagnoses, which must cite collected evidence. Model outputs are evaluated against 21 defined RCA cases.
Hybrid platform on four self-hosted Kubernetes clusters, combining self-hosted inference and hosted models. Legal RAG with full-text and vector search, a validator for cited sources and a grounding step. Controlled agent development with local and larger model reviews. Access platform around open-source OpenBao for agent and workload identities. Governance including DPIA, an AI risk register and consent flows.