I'm Lef. I build production-grade LLM and agent infrastructure for scaling teams.
Stockholm, remote-friendly.
LLM platforms melt under peak load. Agent systems leak data, burn cash, or both. Observability arrives last. Multi-region routing that survives, sandboxed agents you can put in production, cost and latency you can actually see.
What I do
- Custom agentic integrations
- Production bug to Jira ticket to auto-fix PR for simple cases, faster triage on the rest. Agentic interfaces (MCP, webhooks, structured APIs) ship in every product now, but every business has its own workflow edges the off-the-shelf connectors don't fit. Integration that holds when your workflow doesn't match the demo.
- LLM serving reliability
- Multi-region routing, per-service quota isolation, structured rate-limiting, cache-friendly prompting. Stop letting one team's bad day take the platform down.
- Agent infrastructure
- Sandboxed Python execution with read-only permissions and network policies, ground-truth evaluation pipelines, reusable security primitives. Agents you can ship without a security review every Friday.
- Observability and cost
- Prometheus, Grafana, Loki, Alloy. Normalised log levels across Python and TypeScript services. Per-unit LLM cost via workflow-context propagation. The dashboards that make on-call possible.
Selected work
- at Filed Multi-region Gemini load balancer plus per-service project routing. Reduced production rate-limiting from a workflow-blocking issue to a non-event.
- at Filed AI-driven production triage agent: in-cluster LLM service behind a Slack bot, with a sandboxed Python executor reused across other agents.
- at Tink (Visa) Ratchet, internal MLOps framework packaging IaC, CLI, and CI/CD. Data Scientists ship ML to Vertex AI without friction.
- at Northvolt Multi-backend data-access library used daily by 30+ engineers. Enabled production-line incident debugging on the factory floor.
Signals
- NDSML Summit talk: Elevating Your Existing Models to Vertex AI (with Andrew Wu).
- Stockholm MLOps Community talk: Ratchet: Elevating your existing models to Vertex AI (with Erik Pärlstrand), March 2024.
- NordSec 2018: Data Modelling for Predicting Exploits (Springer LNCS 11252).
- Arcturus, browser-based virtual analog synthesizer, ~95% built by Claude Code agents on a custom autonomous loop.