01
Applied AI & GenAI
Building complete AI systems rather than demos — and building the tooling agents consume, rather than only calling APIs that already exist.
Systems
- LLM systems
- RAG pipelines, including hybrid RAG
- LLM orchestration
- AI agents & agentic workflows
- Long-term memory systems
- Evaluation frameworks
- Guardrails
- LLM observability
Protocols & tooling
- MCP servers built, published and operated
- MCP clients
- Coding agents in daily use
- AI tooling design
Inference
- Self-hosted LLMs (Ollama)
- Local GPU inference
- Quantisation trade-offs
- Privacy-first AI
- Cost / latency / quality arbitration
02
Software & systems engineering
One core, many surfaces, one public contract — and the packaging discipline that lets other people actually install what you wrote.
Design
- Software architecture
- Backend engineering
- API design
- Distributed systems
- Concurrency & async
- Domain modelling
- Design for reliability
Delivery
- Python (primary language)
- FastAPI
- REST APIs
- Python SDKs on PyPI
- CLI tooling
- Browser extensions
- Monorepo & validation gates
- Semantic releases
03
Infrastructure, DevOps & platform
The glue that decides whether an AI system survives contact with real users. Every incident I have driven to root cause ended in a pipeline check, an alert or an automation — not a patch.
Platform
- Docker & containerisation
- Proxmox & LXC
- Kubernetes / k3s
- Helm
- Terraform
- Linux
- Networking
Delivery & operations
- CI/CD (Forgejo, GitHub Actions)
- Git-driven releases & multi-arch images
- Monitoring & alerting
- OpenTelemetry, Grafana, Elastic
- SLI/SLO
- Incident response & root-cause analysis
- Backups & recovery
04
Linux & system administration
The deepest layer of this profile, and the one that took longest to start claiming out loud.
Depth
- Arch Linux as daily production machine for ten years
- Custom window manager wrapper around i3
- ~50 distributions installed and broken
- Package management & dependency conflict resolution
- systemd, init, daemon authoring
Scale today
- ~30 Linux systems administered at once
- Debian, Ubuntu, Alpine
- LXC containers, Docker containers, VMs
- Encryption & TPM
- GPU passthrough
05
Data & cloud platforms
Four years of enterprise data platforms in a regulated environment, from ingestion through to governance.
Platforms
- Databricks
- Spark / PySpark
- Delta Lake
- Unity Catalog
- Lakehouse architecture
- SQL
Cloud
- AWS (platform of choice)
- Azure & Power BI (narrow, real)
- Ingestion pipelines
- Streaming
- Data governance & quality
06
Classical machine learning
Where this started, and still the reason the mathematics under a model is never a black box.
Methods
- Time series
- Anomaly detection
- Supervised & unsupervised learning
- Deep learning fundamentals
- Statistics
- Pandas, NumPy
Calibration
Where I draw the line
A profile is only as good as its weakest claim. These are the five places
where the honest sentence is narrower than the one a keyword list would
suggest — stated here so that nobody has to discover them later.
- Kubernetes
-
+ what is true
Hands-on in a personal lab environment — building, deploying and breaking workloads on a self-managed k3s cluster, plus the observability stack on top of it.
− what is not
My production services do not run on it. They run on Docker under Proxmox. I have not operated Kubernetes at enterprise scale, nor as a shared service for other teams.
- Observability
-
+ what is true
OpenTelemetry, Grafana, Elastic and Helm deployed and run on my own cluster, with a first pass at SLIs and SLOs. Installation, configuration, queries, dashboards.
− what is not
Not an observability platform operated for other teams, and not a formalised SLO practice sustained over years.
- GitOps
-
+ what is true
Git as the source of truth for configuration and releases: push- and tag-driven pipelines, semantic releases, multi-arch images, published packages.
− what is not
No reconciliation operator — no Argo CD, no Flux — runs on my cluster. 'GitOps-style delivery', not 'a cluster reconciled from Git'.
- MLflow
-
+ what is true
Governance and supervision of MLflow projects on Databricks with Unity Catalog — I put the conditions in place and oversaw the work.
− what is not
Not deep hands-on practice: experiment tracking, model packaging and registry were not my daily work.
- Azure
-
+ what is true
Real enterprise exposure: Power BI delivery and cluster management on a client platform, as the manager of that work.
− what is not
Narrow — visualisation and operation, not Azure platform engineering. AWS is by a distance the cloud I am most fluent in.