The curriculum

1. Week 716 Nov 2026

   Ingestion and document pipelines

   Design: Invoice and claims processing (insurance)

   • Parse messy documents (PDF, scans, tables) and measure extraction quality
   • Choose between batch, streaming and change data capture for keeping an index fresh
   • Re-index incrementally without downtime or duplicate chunks
2. Week 823 Nov 2026

   Vector, hybrid and graph retrieval

   Design: Semantic product search over 50 million items (e-commerce)

   • Explain HNSW, IVF and quantisation trade-offs and size memory and cost for an index
   • Combine keyword, vector and metadata filters at scale
   • Judge when a knowledge graph (GraphRAG) earns its extra cost
3. Week 930 Nov 2026

   Data quality, contracts and lineage

   Design: Text-to-SQL analytics assistant over governed data (banking)

   • Write a data contract with owners, schema and freshness targets
   • Trace an AI answer back to the source rows and documents it used
   • Classify and protect personal data before it reaches a model
4. Week 107 Dec 2026

   Identity and permission-aware access

   Design: HR and legal assistant with strict access rules (enterprise HR)

   • Compare RBAC, ABAC and relationship-based access for AI retrieval
   • Keep permissions in sync between source systems and the index
   • Give agents scoped, auditable credentials instead of shared keys
5. Week 1114 Dec 2026

   Context engineering and memory

   Design: Sales and CRM agent with long-running account memory (B2B software)

   • Budget a context window across instructions, retrieval, history and tools
   • Design short-term and long-term memory, and decide what to forget
   • Pick a compaction strategy and measure what it loses
6. Week 1221 Dec 2026

   Review: recap, self-test and Learning Radar

1. Week 1328 Dec 2026

   Inference serving: batching, KV cache, quantisation, GPU choice

   Design: Self-hosted model platform for internal teams (technology company)

   • Explain continuous batching, paged KV cache and where each GPU's memory goes
   • Choose quantisation and parallelism settings for a target model and latency
   • Decide between hosted APIs and self-hosted serving with numbers
2. Week 144 Jan 2027

   Caching layers: prompt, semantic and result caching

   Design: High-volume customer FAQ assistant (airline)

   • Lay out prompts so provider and self-hosted prefix caches actually hit
   • Design a semantic cache with a safe similarity threshold and invalidation
   • Measure cache savings and the wrong answers a cache can cause
3. Week 1511 Jan 2027

   Model routing, small and on-device models, cost optimisation

   Design: Call summarisation at one million calls a month (telecom)

   • Design a router or cascade that sends easy requests to cheaper models
   • Judge when a small or on-device model is good enough
   • Use batch and flex tiers and build a cost-per-request model
4. Week 1618 Jan 2027

   Latency engineering: streaming, speculative decoding, parallel tools

   Design: Real-time voice agent (hospitality)

   • Break end-to-end latency into its parts and find the biggest one
   • Apply streaming, speculative decoding and parallel tool calls where they help
   • Set latency budgets per step and enforce them
5. Week 1725 Jan 2027

   Capacity planning and GPU economics

   Design: Code-assistant platform for 5,000 engineers (software)

   • Estimate tokens, GPUs and cost for a forecast load
   • Plan for peaks, regions and failover capacity
   • Compare reserved, on-demand and API pricing over a year
6. Week 181 Feb 2027

   Review: recap, self-test and Learning Radar

1. Week 198 Feb 2027

   AI security and threat modelling

   Design: Customer-facing banking agent with tools (banking)

   • Threat-model an AI system with tools (prompt injection, data exfiltration, excessive agency)
   • Apply layered defences and know which ones are only partial
   • Run red-team tests and turn findings into regression tests
2. Week 2015 Feb 2027

   Reliability patterns: idempotency, queues, backpressure, durable workflows

   Design: Claims-processing agent pipeline (insurance)

   • Make every side effect safe to retry
   • Use queues and backpressure to survive provider outages and rate limits
   • Choose between a workflow engine and a simpler queue
3. Week 2122 Feb 2027

   LLMOps: versioning and CI/CD for prompts and models, canary, shadow

   Design: Release process for a multi-team GenAI platform (retail group)

   • Version prompts, models, tools and eval sets together
   • Release with canary and shadow traffic, and roll back safely
   • Handle a provider model deprecation without an incident
4. Week 221 Mar 2027

   Multi-tenancy, quotas and chargeback

   Design: SaaS product with per-tenant AI features (SaaS)

   • Isolate tenants' data, prompts, caches and limits
   • Enforce quotas and fair sharing under load
   • Attribute cost per tenant and feature
5. Week 238 Mar 2027

   Disaster recovery, multi-region and data residency

   Design: Citizen services assistant under GCC and EU residency rules (government)

   • Set RPO and RTO for each AI component, including indexes and memory
   • Design multi-region serving that respects residency rules
   • Plan for a model provider outage
6. Week 2415 Mar 2027

   Review: recap, self-test and Learning Radar

1. Week 2522 Mar 2027

   Fine-tuning in practice: full fine-tuning, LoRA, QLoRA

   Design: Replace a frontier-model call with a fine-tuned small model for document extraction (insurance)

   • Decide when fine-tuning beats prompting or RAG, with evidence
   • Prepare training data and evaluate the result against the base model
   • Estimate the cost of training and serving a fine-tuned model
2. Week 2629 Mar 2027

   Training data, synthetic data and distillation

   Design: Customer-care model built by distillation (telecom)

   • Curate and clean a training set and measure its quality
   • Generate synthetic data without teaching the model its own mistakes
   • Distil a large model into a small one, and handle data licensing
3. Week 275 Apr 2027

   RL fundamentals for LLMs: policy gradients, PPO, reward models, RLHF

   Design: Brand- and policy-aligned service model with a human feedback loop (retail bank)

   • Explain RLHF end to end, from preference data to policy update
   • Design and validate a reward model
   • Recognise reward hacking and over-optimisation
4. Week 2812 Apr 2027

   Modern post-training: DPO, RL with verifiable rewards, GRPO, agent RL

   Design: Maintenance agent trained on internal tools and work orders (manufacturing)

   • Choose between preference tuning and RL for a goal
   • Build reward functions that can be checked automatically
   • Train an agent in an environment built from real tools
5. Week 2919 Apr 2027

   Building your own model: pretraining, continued pretraining, tokenisers, scaling laws

   Design: Arabic-first or Indic-language enterprise model (national champion company)

   • Judge when building or continuing to pretrain a model is worth it
   • Plan data, tokeniser, compute and evaluation for a training run
   • Use scaling laws to budget model size and tokens
6. Week 3026 Apr 2027

   Review: recap, self-test and Learning Radar

1. Week 313 May 2027

   Planning and reflection

   Design: Research analyst agent (asset management)

   • Compare plan-then-execute, ReAct and reflection loops on cost and reliability
   • Make plans reviewable before they run
   • Stop loops that are not converging
2. Week 3210 May 2027

   Agent-to-agent protocols (A2A) across organisational boundaries

   Design: Procurement agents working across suppliers (logistics)

   • Explain how A2A agent cards, tasks and messages differ from MCP tool calls
   • Design trust, authentication and contracts between agents owned by different organisations
   • Handle partial failure and disputes when a partner's agent is slow, wrong or unavailable
3. Week 3317 May 2027

   Human-in-the-loop and approval limits

   Design: Finance operations agent with spending limits (enterprise finance)

   • Place approvals where risk is, not everywhere
   • Design approval UX that people actually read
   • Measure automation rate against error rate
4. Week 3424 May 2027

   Sandboxing and computer use

   Design: IT operations automation (managed services)

   • Isolate agent code execution and browser or desktop control
   • Limit network, file and credential access per task
   • Decide when computer use beats an API integration
5. Week 3531 May 2027

   Agent identity, registry and governance

   Design: Group-wide AI and agent platform (energy utility)

   • Give every agent an identity, an owner and an audit trail
   • Run an internal registry of agents, tools and their permissions
   • Retire and update agents safely
6. Week 367 Jun 2027

   Review: recap, self-test and Learning Radar

Builds on M6. Week 32 covered protocols between organisations; this module is about building, evaluating and controlling teams of agents inside one system.

1. Week 3714 Jun 2027

   Agent Skills and capability packaging

   Design: Internal skill library for a 2,000-person engineering organisation (software)

   • Decide when a capability should be a skill, a tool, an MCP server or a sub-agent
   • Package a skill so it loads only when needed (progressive disclosure) and test that it triggers
   • Version, review and share a skill library across teams without breaking agents that use it
2. Week 3821 Jun 2027

   Agent frameworks and harnesses: choosing one and keeping an exit

   Design: Standardising one agent framework across business units (bank)

   • Compare LangGraph, OpenAI Agents SDK, Claude Agent SDK, Google ADK, CrewAI and Microsoft's agent framework on state, control flow and deployment
   • Learn what coding-agent harnesses (Claude Code, Codex) show about context, tools and permissions
   • Keep business logic portable so a framework change does not mean a rewrite
3. Week 3928 Jun 2027

   Multi-agent topologies and context isolation

   Design: Due-diligence team of agents (private equity)

   • Decide when several agents beat one agent with more tools, using cost and quality numbers
   • Choose between orchestrator-worker, supervisor, hand-off, parallel fan-out and critic designs
   • Isolate each sub-agent's context and design what it hands back, and how state is shared
4. Week 405 Jul 2027

   Evaluating and tracing agents and multi-agent systems

   Design: Go-live gates for a customer-service agent team (telecom)

   • Evaluate whole trajectories and tool calls, not just final answers
   • Test agents against simulated users and tools, and set go-live gates
   • Trace one request across several agents and find where it went wrong
5. Week 4112 Jul 2027

   Failure modes, budgets and control of multi-agent systems

   Design: Autonomous incident-response agents (cloud provider)

   • Recognise loops, error cascades, conflicting writes and runaway cost between agents
   • Set token, time and spend budgets with circuit breakers and a kill switch
   • Run a multi-agent system in production: alerts, on-call and safe rollback
6. Week 4219 Jul 2027

   Review: recap, self-test and Learning Radar

1. Week 4326 Jul 2027

   Recsys foundations: retrieval then ranking, collaborative filtering, two-tower models

   Design: Homepage recommendations (e-commerce)

   • Explain the two-stage retrieval and ranking architecture
   • Build candidate generation with two-tower embeddings
   • Handle cold start for new users and items
2. Week 442 Aug 2027

   Ranking and features: feature stores, learning to rank, deep ranking models

   Design: News and content feed ranking (media)

   • Design features and keep training and serving consistent
   • Choose a ranking model and loss for the business goal
   • Run a feature store in production
3. Week 459 Aug 2027

   Real-time recommendations: session models, event pipelines, freshness

   Design: Real-time recommendations in a delivery app (logistics and food delivery)

   • Build a streaming feature pipeline
   • React to what a user did seconds ago
   • Meet tight latency budgets at peak
4. Week 4616 Aug 2027

   LLMs and generative recommenders: semantic IDs, LLM ranking, conversational recommendations

   Design: Conversational shopping assistant (retail)

   • Judge where LLMs genuinely improve recommendations
   • Explain semantic IDs and generative retrieval
   • Keep LLM-based ranking within latency and cost limits
5. Week 4723 Aug 2027

   Evaluation and experimentation: offline metrics, A/B tests, interleaving, bandits

   Design: Experimentation platform for a streaming service (media streaming)

   • Design an experiment that measures business impact
   • Avoid popularity feedback loops and unfair exposure
   • Use bandits where A/B tests are too slow
6. Week 4830 Aug 2027

   Review: recap, self-test and Learning Radar

1. Week 496 Sept 2027

   Regulation and governance: EU AI Act, UAE and KSA rules, India DPDP, model risk

   Design: Clinical documentation assistant (healthcare)

   • Classify a system's risk under the EU AI Act and list the duties that follow
   • Map data residency and privacy rules to architecture decisions
   • Set up model risk management that auditors accept
2. Week 5013 Sept 2027

   Integrating with legacy systems: APIs, events, mainframe, replacing RPA

   Design: Core-system modernisation with AI (insurance)

   • Connect AI to systems of record without breaking them
   • Choose between APIs, events and batch for each integration
   • Plan a strangler-style migration away from RPA scripts
3. Week 5120 Sept 2027

   Responsible AI: fairness, explainability, content safety, AI policy

   Design: Credit and hiring decisions with human review (financial services and HR)

   • Measure and reduce unfair outcomes
   • Explain AI decisions to users, auditors and regulators
   • Turn an AI policy into enforceable technical controls
4. Week 5227 Sept 2027

   Multimodal pipelines: vision, speech, video

   Design: Field-service inspection from photos (utilities and infrastructure)

   • Design pipelines for images, audio and video at scale
   • Choose between multimodal models and specialised ones
   • Control the cost of multimodal processing
5. Week 534 Oct 2027

   Platform strategy, build vs buy, multi-cloud

   Design: Group-wide AI platform for a conglomerate (diversified group)

   • Decide what to build, buy or assemble, and when to revisit it
   • Avoid lock-in with gateways and portable data
   • Run AI across more than one cloud without doubling the work
6. Week 5411 Oct 2027

   Review: recap, self-test and Learning Radar

Each week is one full reference architecture. Monday revisits the building blocks it needs; Thursday is a full mock interview round on it.

1. Week 5518 Oct 2027

   Capstone: enterprise knowledge platform (M1, M2)

   Design: Enterprise knowledge platform (professional services)

   • Combine retrieval, evals, ingestion, permissions and memory into one design
   • Defend every major decision with an ADR
   • Present it in 10 minutes to a mixed audience
2. Week 5625 Oct 2027

   Capstone: contact-centre AI, voice and text (M3)

   Design: Contact-centre AI (telecom)

   • Design for latency, cost and peak load together
   • Plan capacity and cost for a year
   • Present trade-offs to a CTO and a CFO
3. Week 571 Nov 2027

   Capstone: regulated banking agent platform (M4, M6, M7, M9)

   Design: Regulated banking agent platform (banking)

   • Combine security, multi-agent coordination, approvals and regulation in one design
   • Pass a mock architecture review
   • Write the risk register and operating model
4. Week 588 Nov 2027

   Capstone: own-model platform (M5)

   Design: Domain model platform (healthcare or legal)

   • Plan fine-tuning, alignment, serving and evaluation for a domain model
   • Justify the investment against API alternatives
   • Design the retraining loop
5. Week 5915 Nov 2027

   Capstone: personalised commerce (M8)

   Design: Personalised commerce platform (retail)

   • Combine recommendations with an LLM assistant
   • Design the experiment that proves value
   • Plan real-time features and serving
6. Week 6022 Nov 2027

   Review: recap, self-test and Learning Radar

Tuesday uses a study, survey or well-known postmortem instead of a paper; Saturday produces an artefact (ADR, roadmap, business case) instead of code.

1. Week 6129 Nov 2027

   Running design reviews and writing ADRs

   Design: Architecture review board for AI projects (enterprise)

   • Run a design review that finds real problems
   • Write ADRs people read and follow
   • Review someone else's design constructively
2. Week 626 Dec 2027

   Cost ownership and FinOps for AI

   Design: AI FinOps for a multi-product company (SaaS)

   • Own an AI budget and forecast it
   • Set unit economics targets and track them
   • Cut cost without cutting quality
3. Week 6313 Dec 2027

   Business cases and stakeholder management

   Design: Business case for an AI programme (manufacturing)

   • Write a business case with value, cost and risk
   • Map stakeholders and what each needs to say yes
   • Handle a sceptical executive
4. Week 6420 Dec 2027

   Technical roadmaps and platform adoption

   Design: Roadmap for an internal AI platform (retail group)

   • Build a 12-month roadmap with clear bets
   • Drive adoption of a platform across teams
   • Measure whether the platform is working
5. Week 6527 Dec 2027

   Architecture governance, mentoring and delegation

   Design: AI centre of excellence operating model (enterprise)

   • Set standards without becoming a bottleneck
   • Mentor engineers into architects
   • Delegate design work and keep quality
6. Week 663 Jan 2028

   Review: recap, self-test and Learning Radar