Pricing
AI Platform Engineering
Phase 5Weeks 12-148 machine-verified checks · 1 tutorial pack

Serving, Gateways and Cost

The platform half: run models like production software.

Why this phase exists

This is where platform engineering and AI meet, and where most AI engineers stop being useful. Routing, budgets, fallbacks, GPU economics, and cost attribution are the difference between a demo and a service your company can afford to run.

What you learn

  • Gateways: routing, budgets, fallbacks, per-tenant virtual keys
  • Self-hosted serving (vLLM, KServe-class) and when it actually beats hosted APIs
  • GPU basics for platform people: VRAM math, KV cache, continuous batching, cold starts
  • Autoscaling model workloads without paying for idle accelerators
  • Cost observability: per request, per tenant, per feature
  • Delivered on AWS, GCP, Azure, on-prem, and cloud-agnostic Kubernetes
Outcome: You run the model plane: routed, budgeted, observable, on any cloud.

Tools you will actually touch

LiteLLMvLLMKServeKubernetesPrometheus + GrafanaDCGM exporter

The assignment pack

Scenario

Three product teams share one AI budget and keep blaming each other for the bill. Stand up a model plane that routes their traffic, enforces per-team budgets, fails over when a provider degrades, and attributes every cent to a team and a feature.

Definition of done
  • Gateway in front of all model traffic with per-tenant virtual keys
  • Enforced budget that actually stops spend, not just alerts on it
  • Automatic fallback proven by degrading the primary provider
  • Cost dashboard broken down by tenant and feature
  • A documented decision on hosted versus self-hosted with the numbers behind it
How it is graded
Fail

A shared API key and a monthly invoice surprise.

Pass

Gateway with per-tenant keys, enforced budgets, working fallback, cost dashboard.

Strong pass

Self-hosted path benchmarked against hosted with real throughput and cost figures, and autoscaling that does not pay for idle GPUs.

The tutorial pack

Stand up a model plane

  1. 1Put a gateway in front of one provider and route all traffic through it
  2. 2Issue per-tenant virtual keys and set hard budgets
  3. 3Add a second provider and configure fallback rules
  4. 4Kill the primary provider and prove failover works
  5. 5Deploy a self-hosted model and benchmark tokens per second against the hosted path
  6. 6Export cost and token metrics; build the per-tenant dashboard
The failure you will hit on purpose

Size a model so its KV cache exceeds available VRAM at your target concurrency, then watch it OOM under load. Do the VRAM math afterwards and resize. GPU economics stop being abstract at that point.

What you measure at the end
  • Cost per 1,000 requests, hosted versus self-hosted
  • p95 latency under target concurrency
  • Budget enforcement: spend stops at the ceiling
  • Failover time when the primary provider degrades

Per-target lab variants

Same outcome, five stacks. You train on the one you will actually run.

AWSEKS + Bedrock or vLLM on g5/p4d nodes, cost allocation tags into Cost Explorer
GCPGKE + Vertex or vLLM on L4/A100, billing export into BigQuery for attribution
AzureAKS + Azure OpenAI, Cost Management with per-tag chargeback
On-premKubernetes + vLLM on your own GPUs, DCGM exporter for utilisation and cost modelling
AgnosticAny conformant K8s + LiteLLM + Prometheus, portable and no managed lock-in

You must be able to answer

AI use is mandatory on every assignment. The integrity mechanism is defend-your-work, so these are the questions that decide whether the work counts.

  • At what request volume does self-hosting beat the API, and show your math.
  • A tenant hits their budget mid-request. What does the user see?
  • How much VRAM does your deployment need at 20 concurrent requests?
  • Your provider has a partial outage, 30 percent of calls fail. What happens?

Want to be walked through this phase?

The packs are delivered with live teaching, reviewed assignments, and a defence.

Train with me