The whole AI Architect course on one page, for someone who has read it. This page is a map, not new content: every rule, formula and number below is copied from the lesson in the last column, so open that section to see the reasoning, the assumptions and the date a fact was checked. Page numbers are the course order.
The course on one figure
Figure 1. The reference system, with the page that teaches each part
How to read this: each box carries the pages that teach it, and the figure uses the course's component names. Page 2 (requirements, ADRs and design reviews) applies to the whole design, page 3 teaches how to draw views like this one, page 11 covers the cost of the calls through the Model gateway, page 14 covers the LLM provider as a managed cloud service, and page 15 applies the course in two full designs.
Which page when
Each symptom is the first column of a failure table in the lesson. Follow the link to see what the system should do.
| Symptom | Read |
|---|---|
| A requirement exists only in conversation | Page 2, failure scenarios |
| Only the happy path is drawn | Page 3, failure scenarios |
| A cache returns another tenant's or an old-policy answer | Page 4, failure scenarios |
| A filtered query returns too few results | Page 5, failure scenarios |
| The loop does not terminate | Page 6, failure scenarios |
| A poison message blocks a partition | Page 7, failure scenarios |
| The aggregate is green while a slice regressed | Page 8, failure scenarios |
| The canary result is inconclusive | Page 9, failure scenarios |
| Provider returns 429 for a minute | Page 10, failure scenarios |
| The forecast is off by three times | Page 11, failure scenarios |
| Injected text in a retrieved document asks for a send | Page 12, failure scenarios |
| A system in production is not in the registry | Page 13, failure scenarios |
| Inference is routed outside the allowed region | Page 14, failure scenarios |
| You want to see two complete designs and test whether the whole course holds together | Page 15, capstone |
Part 1: Design the system (pages 2 to 8)
Each table lists decision rules (If, then), formulas and numbers. Left column: the situation, formula or calculation. Middle column: the response or the result. Right column: the section that teaches it.
Page 2: Requirements, ADRs and design reviews
AI Architect Mindset: Requirements, ADRs, Design Reviews
| Remember | Then or result | Read |
|---|---|---|
| If A requirement has no number and no measurement point | Do not build against it | Decision framework |
| If A review ends without a recorded outcome | Treat the design as not approved | Decision framework |
| Formula p ± 1.96 × √(p(1 − p) / n) | only the lower bound can demonstrate a bar | Worked example |
| Calc (1 − 0.995) × 400,000 | 2,000 failed requests in the window | Worked example |
Page 3: Diagramming AI systems
Diagramming AI Systems: C4, Sequence, State, ER
| Remember | Then or result | Read |
|---|---|---|
| If 3 or more participants, or any async step | Draw a sequence diagram | Decision framework |
| If Data crosses to a party you do not control, or model output or retrieved text can trigger an action | Mark a trust boundary and name its control | Decision framework |
| Number About 15 nodes at most | Over that, split into an overview and zoom-ins | The diagram checklist |
| Calc 3 x 3 x 3 = 27 attempts at the ticketing system | the Orchestrator is the only layer that retries | Worked example |
Page 4: Models, routing and inference
Models, Routing and Inference for AI Architects
| Remember | Then or result | Read |
|---|---|---|
| If The request type is known before the call | Route with rules first | Decision framework |
| If An answer depends on user, tenant, time or changing data | No semantic cache; exact caches keyed on all inputs | Decision framework |
| Formula Bytes per token = 2 (key and value) x layers x KV heads x head dimension x bytes per value | Context length becomes memory | Attention and the KV cache |
| Formula With the large model at cost 1 and the small at 0.1, a cascade costs 0.1 + f for escalation rate f | It stops paying at 0.9 | Routing policies |
| Formula cached cost for n requests = 1.25 + 0.1 (n - 1) prefix-lengths | Anthropic 5-minute multipliers, as of 2026-10-22 | Prompt caching and prefix stability |
Page 5: Data and retrieval
Data and Retrieval Architecture for AI Systems
| Remember | Then or result | Read |
|---|---|---|
| If Callers see part of the corpus | Filter from verified identity inside the search | Decision framework |
| If Queries contain identifiers or defined terms | Add a lexical index and fuse with RRF | Decision framework |
| Formula RRFscore(d) = sum over rankings r of 1 / (k + r(d)) | with k = 60 | Reciprocal rank fusion |
| Calc 40 × (1/500) | 0.08 allowed results per top-40 | Worked example |
Page 6: Agents, tools and MCP
Agents, Tools and MCP: Architecture for AI Architects
| Remember | Then or result | Read |
|---|---|---|
| If A tool is irreversible | Key, policy decision, approval above a limit | Decision framework |
| If A loop can run long | Cap calls, tokens and time from measured runs | Decision framework |
| Formula Total(N) = sum of I(k) for k = 1..N = N * (S + U) + (a + r) * N * (N - 1) / 2 | the total grows with the square of the steps | Context growth and compaction |
| Number The loop costs 4.7 to 9.3 times the input tokens | 2.7 to 4.0 times the latency | Worked example |
Page 7: Multi-agent and event-driven designs
Multi-Agent and Event-Driven AI Architecture
| Remember | Then or result | Read |
|---|---|---|
| If Delivery can repeat | Commit after processing; idempotent effect keyed by event id | Decision framework |
| If State change and event must agree | One local transaction through an outbox | Decision framework |
| Rule Retention must cover the longest period one event can be redelivered or replayed | Keys are kept 24 hours and the topic keeps events 7 days | Idempotency keys |
| Formula in-flight work = arrival rate x time in the system | Little's law | Worked example |
| Number capacity is partitions divided by step time | 4 events per second of headroom turns a 3-minute outage into a 15-minute recovery | Worked example |
Page 8: Evaluation
Evaluating AI Systems: Metrics, Judges and Gates
| Remember | Then or result | Read |
|---|---|---|
| If A gate blocks merges | Confidence bound; state the detectable drop | Decision framework |
| If A judge decides the release | Calibrate; report kappa and failure recall | Decision framework |
| Formula kappa = (P(A) - P(E)) / (1 - P(E)) | Cohen's kappa asks how much better than chance the agreement is | Calibrating a judge |
| Number With n = 400, a one-sided 95% lower bound and a 0.90 threshold it needs 370 passes (92.5%) | A system at 0.92 fails this 0.90 proof gate 60% of the time at n = 400 | Confidence intervals and sample sizes |
| Number about 3.6 points at 400 items | about 16 for a 30-item slice | At a glance |
Part 2: Operate, secure and govern (pages 9 to 15)
Page 9: LLMOps and observability
LLMOps and Observability for AI Systems
| Remember | Then or result | Read |
|---|---|---|
| If Any answer-shaping part changes | Mint a release ID and stamp it on every span | Decision framework |
| If A change has side-effecting tools | Shadow with stubbed tools and canary behind approval | Decision framework |
| Formula ((1.645 x sqrt(0.92 x 0.08) + 0.8416 x sqrt(0.88 x 0.12)) / 0.04)^2 | 323.8, so 324 items | Worked example |
| Number halving the drop roughly quadruples the sample | about 1,221 (1,271 with the correction) | Worked example |
Page 10: Reliability and scaling
Reliability and Scaling for AI Systems
| Remember | Then or result | Read |
|---|---|---|
| If Several layers could retry | One owner per failure class; SDK retries to 0 | Decision framework |
| If Many clients retry on one signal | Full or decorrelated jitter and an attempt limit | Decision framework |
| Formula L = λ x W | L / calls per worker / target utilisation | Capacity and Little's law |
| Formula v = min(cap, base x 2^n) | uniform(0, v) | Backoff with jitter |
| Number Refill is the limit divided by 60 | 300 per minute is 5 per second | Rate limiting algorithms |
| Number SRE book: retry only while the retry ratio is below 10% | up to three attempts, then let the failure bubble up | One retry owner |
Page 11: Cost and unit economics
Cost and Unit Economics for AI Systems
| Remember | Then or result | Read |
|---|---|---|
| If Per-task cost is not measured | Instrument at the Model gateway before changing anything | Decision framework |
| If Output tokens dominate | Shorten format and max_tokens first |
Decision framework |
| Formula cost per successful task = (model + retrieval + tools + retries + review + infrastructure) / share of tasks that succeed | A model one third the cost per attempt (1 CU against 3) that succeeds 25% of the time costs 1 / 0.25 = 4 CU per success | Cost per successful task |
| Formula h* = (w - 1) / (w - r) = 0.25 / 1.15 = 21.7% for w = 1.25 and r = 0.1 | Caching wins when h r + (1 - h) w < 1 | Cache hit-rate math |
| Formula Effective ratio for the batchable work = (1 - f) b + f = b + (1 - b) f | let b be the batch cost ratio (0.5) | Batch versus real-time |
| Formula Break-even escalation rate f* = 1 - (c_s + c_g) / c_l | Cascade cost per request = c_s + c_g + f x c_l | Cascade break-even |
Page 12: Security and threat modeling
Security and Threat Modeling for AI Systems
| Remember | Then or result | Read |
|---|---|---|
| If The model's output can trigger a tool | Authority comes from the session and a scoped credential, never from a model-written value | Decision framework |
| If An action is irreversible or external | Require approval, show the exact content, restrict destinations | Decision framework |
| Calc 2,000 x 0.15 | 300 a day | Worked example |
| Number approving everything costs 6.7 times as much reviewer time | 40,000 s against 6,000 s | Worked example |
| Number the one-sided 95% Wilson lower bound on the block rate is 99.1% | It says nothing about attacks outside the suite | Worked example |
Page 13: Governance and responsible AI
Governance and Responsible AI in Practice
| Remember | Then or result | Read |
|---|---|---|
| If A team wants to ship an AI feature | Register it with its intended purpose before launch | Decision framework |
| If Decision records hold personal data | Keep ids and hashes in the audit store, content elsewhere | Decision framework |
| Rule The risk tier follows the intended use, not the model | record the intended purpose per system | At a glance |
| Calc 120 MB x 365 | 43.8 GB | Worked example |
| Calc (40 + 6.13) / 6 | 7.7, so 8 people | Worked example |
The lesson quotes the EU AI Act text and its application dates with the date they were checked. Read the lesson's section on where the text stands before relying on any date, and treat nothing here as legal advice.
Page 14: Cloud platforms for AI
Cloud Platforms for AI: Networking, Identity, Keys, Logs
| Remember | Then or result | Read |
|---|---|---|
| If A workload holds a long-lived key | Move to workload identity | Decision framework |
| If The model endpoint is public | Add a private endpoint and close the public path | Decision framework |
| Calc Re-encrypt 1,000 x 50 MB = 50,000 MB (50 GB) | Re-wrap 1,000 data keys of 32 bytes = 32,000 bytes of key material | Customer-managed keys and envelope encryption |
| Calc 800 MB x 365 = 292,000 MB = 292 GB | content logs are 16 times larger | Logging and audit trails |
| Calc 40,000 x 365 | 14,600,000 events a year, each checked | Worked example |
The page names the provider products for each job and dates every provider detail with "as of". Read the page before relying on one, and treat the numbers above as its worked assumptions, not as provider limits.
Page 15: Capstone
Page 15, the capstone holds two full designs and the readiness quiz. Use it to check that you can apply every page above in one design. This sheet carries no rule from it.
Decision frameworks and failure tables
Every lesson from page 2 to page 14 ends its teaching with a decision framework (If, then, because, revisit when) and a failure table (failure, what the system does, why). Use them as checklists in a design review.
The same questions, one per line and grouped by page, are in the design-review checklist, which you can also download. The ADR template, which you can also download, is on the same page.
Last verified: 2026-10-23