Sovereign AI
Deployment.
Five days. From hardware spec to a fully air-gapped, monitored, governed LLM platform. The course we wish existed when we built our first one.
Who it's for
Senior infrastructure and platform engineers.
SREs, platform leads, and ML platform engineers responsible for running AI workloads inside the perimeter.
Five-day syllabus
From bare metal to governed platform.
DAY01
Architecture for sovereignty
- What 'sovereign' actually means: data residency, key custody, supply chain
- Hardware landscape: NVIDIA H100/H200, AMD MI300, Habana — trade-offs
- Networking: NVLink, InfiniBand, RoCE, and the cost of getting it wrong
- Storage tiers: model weights, KV cache, embeddings, training checkpoints
LabSpec a 70B-class production cluster: GPU count, network topology, storage layout, and projected throughput. Defend trade-offs to the cohort.
DAY02
Inference engines & quantisation
- vLLM, TGI, TensorRT-LLM, llama.cpp/Ollama — pick the right tool
- Quantisation: GPTQ, AWQ, GGUF, FP8 — accuracy vs throughput
- Continuous batching, PagedAttention, and KV cache management
- Multi-tenant serving, request routing, and per-tenant rate limits
LabDeploy Llama-3 70B on the lab cluster with vLLM. Measure throughput at multiple concurrencies. Apply AWQ quantisation; quantify accuracy delta on a held-out set.
DAY03
Adapting models on-prem
- LoRA, QLoRA, and adapter merging — when fine-tuning is the right answer
- Preference tuning: DPO, ORPO — practical recipes
- Data privacy in training: synthetic data, differential privacy basics
- Evaluation that respects your domain (not just MMLU)
LabFine-tune a domain LoRA adapter on a provided regulated-industry dataset. Compare base vs adapter on the domain eval set; document gains and regressions.
DAY04
Production hardening
- Air-gapped install bundles: registries, model artefact signing, SBOM
- GPU monitoring, NVML, DCGM, and saturating-the-fleet detection
- Secrets, KMS/HSM integration, and API key rotation
- RBAC, audit logging, and the regulator-friendly trail
LabBuild a fully reproducible air-gapped install bundle for the platform stack: container images, model weights, configuration, and a one-command bootstrap.
DAY05
Compliance, operations, and capstone
- EU AI Act risk classification — applied to your real systems
- GDPR data residency and DPIA essentials for AI workloads
- ISO/IEC 42001 (AI management system) — what it asks for, in practice
- Model incident response: drift, prompt-injection events, leakage
CapstonePresent your organisation's sovereign AI roadmap — one slide each for hardware, serving stack, governance posture, and 90-day next steps.
Lead instructor
Shady Ali
Shady Ali
Co-Director · Sadiqoon Technologies
Has architected and operated sovereign AI deployments for institutions where data residency is non-negotiable. Teaches the patterns he has had to defend in front of regulators.
What's included
Inclusions.
- 35 hours of live instructor-led tuition
- Dedicated GPU lab cluster with multiple model sizes
- Reference architecture diagrams and Terraform/Helm templates
- Printed bilingual workbook + private code repository
- Daily lunch & refreshments; cohort dinner Day 4
- UK visa support letter on confirmed registration
- 60 days of post-course Q&A access (extended for AI-201)
- Sadiqoon Institute Certificate of Completion
Register
Reserve your seat.
Submit the form below and we will respond within two working days.