Capstone — Mini GPU Orchestration Platform

Phase: 12 — Leadership & Capstone | Difficulty: ⭐⭐⭐⭐⭐ | Time: 1.5 weeks Language: Python (stdlib) | Hardware: none

Concept primer: ../WARMUP.md Ch. 8.

Run

python platform.py        # boots the platform + runs all integration scenarios

0. The mission

Compose the prior phases into one working control plane and prove the integration with scenarios. The platform is the portfolio piece — it shows you can build the system the JD describes, not just its parts.

request -> [License gate, P11] -> [Admission control, P06/07] -> [HAL routing, P09]
        -> [Paged-KV continuous-batching serving, P07] -> [Metrics + SLOs, P10]

1. What the scenarios prove

ScenarioIntegratesAsserts
1 healthy servingP07 continuous batchingall 20 requests served
2 expired licenseP11 license gateserving refused
3 capability routingP09 HALFP8 → FP8-capable backend; refused if none
4 backend failoverP09 + P10 healthFP16 fails over to a healthy backend
5 overloadP06/07 admissionKV admission protects goodput; nothing OOM-crashes
6 observabilityP10valid Prometheus export

All pass on python platform.py.

2. Reading order (platform.py)

  1. License (P11), HAL/Backend (P09), KVBlockPool (P07), Metrics (P10) — the subsystems.
  2. Platform.serve — the integrated request path (the heart): license gate → admission → routing → batched generation → metrics. Trace one request through it.
  3. The six scenarios in main.

3. The deliverables (write these)

  1. DESIGN-DOC.md — the architecture: the subsystems, the abstraction boundaries (HAL ABI, runtime contract, control-plane API — the seams), the build-vs-buy decisions, the failure modes, what changes at real scale. A template is provided.
  2. RUNBOOK.md — operational procedures for the platform's failure modes (compose the Phase 10 runbooks). Template provided.
  3. The four leadership artifacts in ../leadership/ — strategy, org, OEM, standards.

4. Extensions

  1. Multi-node (P08): TP-shard a model across simulated nodes; add the collective cost to the serving path.
  2. TypeScript control-plane CLI (the JD's JS/TS-for-tooling line): a pctl CLI that talks to the platform's control API (status, license, route).
  3. Real backend: wire the Phase 09 C HAL (CUDA backend if hardware exists) under this control plane.
  4. Attestation gate (P11): require a valid attestation before serving with decrypted weights.

5. What this capstone proves about you

You can integrate — compose a HAL, scheduler, serving engine, licensing, and observability into one coherent platform with sensible abstraction boundaries and honest failure handling. Integration is harder than any single phase and is exactly what a Head of Engineering is hired to lead. With the design doc and leadership artifacts, this is the portfolio that proves you can do the role.