AI
What I do with AI, day to day:
- build
- LLM features in production: persona-driven conversations, prompt and reasoning tuning, a second model as guardrail, retries and fallbacks when a provider goes down.
- models
- Open models through OpenRouter, picked per job on quality, speed and cost: DeepSeek, Qwen and GLM.
- code
- Frontier models for the code itself: mostly Claude Opus in Claude Code for writing it, and GPT-6 Astra in Codex for reviewing it. A local Qwen on my own hardware for cheap experiments.
- delegate
- A team of coding agents: every pull request gets paired bug and security reviews, plus scheduled bug hunts, architecture reviews and weekly summaries.
- extend
- Custom agent skills, plugins and MCP integrations to give agents access to cluster, metrics and data.
I review and own what ships: the agents do the legwork, I do the judgement.
About
Hi, I'm Jeroen. I'm a freelance software/devops engineer (Wellner IT Consultancy), building things for the internet since 1996.
These days that means shipping LLM features into production products, picking and evaluating the models behind them, and running a team of AI coding agents alongside my own work.
I also run what I build: when production misbehaves (slow queries, failing alerts, mail that doesn't arrive) I'm usually the one who finds out why, and I make sure releases ship safely.
AI
What I do with AI, day to day:
- build
- LLM features in production: persona-driven conversations, prompt and reasoning tuning, a second model as guardrail, retries and fallbacks when a provider goes down.
- models
- Open models through OpenRouter, picked per job on quality, speed and cost: DeepSeek, Qwen and GLM.
- code
- Frontier models for the code itself: mostly Claude Opus in Claude Code for writing it, and GPT-6 Astra in Codex for reviewing it. A local Qwen on my own hardware for cheap experiments.
- delegate
- A team of coding agents: every pull request gets paired bug and security reviews, plus scheduled bug hunts, architecture reviews and weekly summaries.
- extend
- Custom agent skills, plugins and MCP integrations to give agents access to cluster, metrics and data.
I review and own what ships: the agents do the legwork, I do the judgement.
Stack
Currently working with (updated November 2026):
- build
- TypeScript · Node.js · Fastify · Vite · PostgreSQL · S3 storage
- ship
- Docker · GitHub Actions · Flux GitOps · SOPS
- run
- Kubernetes (k3s, DOKS) · Traefik · cert-manager · Tailscale
- observe
- Prometheus · Loki · Alloy · Grafana
- tune
- PostgreSQL query & index tuning · connection leaks · p95 latency · replicas
- mail
- deliverability · SPF/DKIM/DMARC · bounce & suppression handling · routing & webhooks
- ai
- OpenRouter · DeepSeek · GLM · Qwen · Grok · local Qwen · MCP
- agents
- Claude Code + Opus (coding) · Codex + GPT-6 Astra (review) · bb (orchestration, plugins, skills)
This site ships the same way: tag, GitHub Actions builds the image, Flux rolls it out.