How this track works
Each module is the theory — why the architecture matters, what to measure, what to expect. Each lab is the practice — exact procedure, recording sheet, observation prompts. The lab is where the module's claims meet your hardware.
You can read the modules in any order, but the track is designed to compound: foundations precede infrastructure, infrastructure precedes inference, inference precedes governance. The eighth module is where everything closes the loop.
Hardware you'll need
Minimum for the first half: one Apple Silicon Mac (M1 or newer, 16GB+ unified memory). For the cluster path: three to six Intel Mac Minis used. For the inference substrate: a Mac Studio or M-series Mac Mini Pro. See the Forge for the full hardware picture →
The eight modules
01
Foundations — pick the hardware
Why Mac Mini + Mac Studio. Power/thermal envelope. Used-market sourcing. The unified-memory math that determines what you can run locally.
Published Paired with L-001
02
First node — Ubuntu on Intel Mac Mini, K3s single-node bring-up
Linux on the Mini that was never advertised as a server. K3s in 15 minutes. The control plane on a shelf.
Planned
03
Multi-node — adding workers, role specialization
Worker-01 through worker-05. Why service specialization beats a homogeneous cluster, even at six nodes.
Planned
04
GitOps — ArgoCD, Harbor, and the artifact pipeline
Declarative cluster state from a Git repo. Why models and containers can share a registry. The promotion pattern.
Planned
05
Observability — Prometheus, Loki, Tempo on a dedicated node
The ISR model applied to your own cluster. Why a dedicated telemetry node is non-negotiable.
Planned
06
Inference — Ollama and MLX on M-series, model selection, the unified-memory math
What 24GB, 64GB, and 192GB actually let you do. Daily-driver models. The 70B test case. Quantization tradeoffs.
Planned Paired with L-002
07
Governance — installing ClawLaw, the policy/enforcement split
The pre-commit gate. Why policy management lives in the cluster and enforcement on the boundary. ALLOW · DENY · ESCALATE wired end to end.
Planned Paired with L-003
08
Closing the loop — full pipeline from commit to evidence record
Training job → registry → inference deployment → metrics → governance gate → evidence. The MLOps lifecycle visible end to end.
Planned
Two audiences. Same track.
If you are building a home lab
The internet has a lot of buy this and run this videos. They make running a local model look like a single command — and a single command is where they end. This track is built on the assumption that you want to actually understand what's happening under your hands. The lab pairs give you something to measure on your hardware so you can build an empirical picture of the architecture, not just an opinion of it.
If you are evaluating practitioners
Each module is a teaching artefact. Each paired lab is a measurement protocol. Together, they demonstrate something a CV cannot: that the practitioner can explain why something works, hand someone a way to verify it, and write up the findings honestly. That is the discipline platform teams need.
Where this track sits
- Silicon hub — the argument for governed Apple Silicon infrastructure
- The Forge — the operational platform these modules build toward
- MLOps architecture — the lifecycle these modules teach in pieces
- Decisions — the architecture choices that shaped the Forge
- The lab — the rigorous experiments each module pairs with