Containers for GenAI & Agentic AI

SCHOOL OF DEVOPS & AI · MLOps / LLMOps

The Open-Source Way

Gourav Shah · 2 Full Days · Intermediate → Advanced · Hands-on

One real system — built one step per module

You don't learn tools in isolation. You ship something real, one block at a time.

From a bare model call to a shipped multi-agent platform — nothing thrown away along the way.

Docker Desktop is now paid for larger orgs. So we build on the open standard instead.

OCI + the Compose Spec run identically on every runtime. Docker becomes one option, never a requirement.

By the end of two days — as a DevOps or app/AI developer.

Everything containerized, portable, and reproducible — using only open-source tooling.

A clean GenAI app and a real agentic system — connected at the tool boundary.

Use Case B uses Use Case A as one of its tools — so skills compound instead of forming one giant tangle.

Introduced the way teams actually adopt AI — each pattern when it earns its keep.

You learn not just how each pattern works, but when it's the right one to reach for.

Each module adds exactly one rung to the same growing system.

M3B (LoRA/QLoRA fine-tuning) is an optional bonus rung for cohorts that want model customization.

Apple Silicon can't expose its Metal GPU to a container. So don't fight it — wire around it.

On Mac: serve the model natively, containerize everything else. On Windows + NVIDIA the server can live in a container.

Think of it as a wall socket — swap the power station behind it and your appliance never notices.

Every serving lab ships behind the same endpoint — so application and agent code stay untouched when the backend swaps.

A non-negotiable design rule — engineered lean on purpose, not by luck.

Target per lab: ≈ 4–6 GB RAM and 2–3 containers. Heavy GPU work stays opt-in with CPU / cloud fallbacks.

A tidy stack: for each job, the tool we actually use in the labs.

Open standards throughout — swap any tool for its peer and the pattern still holds.

One growing compose.yaml, hand-authored service by service across the modules.

By the end, one file tells the whole story — and you understand every line because you wrote it.

Two days, two halves — serve & package, then reason & ship.

Optional M3B (LoRA/QLoRA) can extend the program toward 2.5 days for customization-focused cohorts.

All free and open source — a laptop and a bit of container fluency.

Windows + NVIDIA unlocks the full local vLLM-GPU lab; without it, the CPU track covers the learning.

LET'S BUILD

Container-native, not Docker-native — the open standard is the through-line.

Start with Setup → Prerequisites, then Module 1. · Gourav Shah · School of DevOps & AI