The atlas graph
Click any concept to see what to learn first, what it unlocks, and the portfolio deliverable for its stage.
The critical path — longest prerequisite chain in the graph
1Turn business needs into AI solutions
The most important theme: find the problem worth solving, choose the simplest design that solves it, prove the value, and win the work.
| Atlas stage | Specific concepts (atlas nodes) | What you'd actually do with a client |
|---|---|---|
| 1 · AI literacy | AI capability map, limits and failure modes, model and vendor landscape | Tell a bank's leadership, in plain words, what AI can do for them this year and what it can't |
| 2 · Discovery & framing | Outcome framing, workflow discovery, feasibility and data readiness, requirements and SLAs, success metrics | Run discovery workshops on a process such as loan applications, and turn "we want AI" into a signed discovery brief |
| 3 · Solution patterns | Pattern selection, build vs buy vs partner, human handoff design, reference architecture | Recommend RAG plus a reviewed workflow instead of a fully autonomous agent, and explain why |
| 4 · Business case | Token and inference cost, ROI and value model, total cost of ownership, pricing | Show the cost per processed application against today's manual cost, with a payback period |
| 5 · Strategy & roadmap | Assumption and risk mapping, use-case portfolio, roadmap, AI operating model | Turn ten requested use cases into a phased 12-month roadmap the CIO can fund |
| 6 · Prototype & pre-sales | Smallest testable slice, specifications for agents, rapid prototype, PoC with exit criteria, proposal | Build a two-week PoC that answers the riskiest question, then pitch it together with the business case |
2Build AI agents and automate business workflows
The other most important theme: the building blocks, the agent patterns, the protocols that connect agents to company systems, and the production shape that makes them safe to run.
| Atlas stage | Specific concepts (atlas nodes) | What you'd actually do with a client |
|---|---|---|
| 7 · LLM building blocks | Prompting and task decomposition, structured output, function calling, RAG, context engineering | Build a policy assistant grounded in the bank's documents that cites its sources and returns structured answers |
| 8 · Agent patterns | Agent loop, reflection, planning, memory, multi-agent orchestration | Design a triage agent that classifies customer emails, drafts replies, and sends hard cases to a reviewer |
| 9 · Frameworks & protocols | Agent frameworks and SDKs, MCP, A2A, enterprise integration | Expose the CRM and ticketing system to agents through MCP servers, with read and write tools kept separate |
| 10 · Workflow automation | Production runtimes, human-in-the-loop approvals, document workflows, agent observability, cost governors | Run invoice processing on an event trigger, with human approval above a threshold, tracing, and a monthly spend cap |
3Evaluate, govern, deliver and lead
The supporting stages that decide whether a solution is trusted, allowed to ship, delivered on time and actually adopted.
| Atlas stage | Specific concepts (atlas nodes) | What you'd actually do with a client |
|---|---|---|
| 11 · Evaluation & quality | Eval datasets, LLM-as-judge, error analysis, A/B testing, evals as release gates | Prove to the client's risk team, with evidence, how often the assistant answers policy questions correctly |
| 12 · Governance & risk | Data classification, guardrails and PII masking, prompt-injection defense, least privilege, compliance, audit | Produce the control matrix the bank's compliance and audit teams need before go-live |
| 13 · Project delivery | Decision records, pilot-to-production rollout, SLA operations and runbooks, feedback loop | Take the PoC to production through gated rollout, with an SLA, a runbook and a decision log |
| 14 · Leading AI teams | AI team design and roles, change management and adoption, learning culture | Plan the roles, training and adoption measures so branch staff actually use the new system |
4An AI Engineering from Scratch route
Start here: it is the closest fit for this atlas. 27 lessons from AI Engineering from Scratch, an open, independent community curriculum. Lessons tagged CCAR-P belong to its Claude Certified Architect – Professional study track, and their links open inside that track. Mappings come from each lesson's learning objectives.
5A Stanford route
23 courses in five phases, chosen from a scan of 1,204 AI-related courses across every Stanford department in the 2026–27 catalog, not only Computer Science. Business courses come from the Graduate School of Business (GSB), Management Science & Engineering and Electrical Engineering. Mappings come from the official course descriptions. Each card links to the course's public syllabus or course site where one exists (11 of 23). The other 12 have no public syllabus, so those cards keep the catalog link.
6A DeepLearning.AI route For technical work
Technical track. These courses teach the building skills: prompting, RAG, agent frameworks, protocols, evaluation and guardrails. They say little about discovery, business cases or leading change, so take them alongside routes 4 and 5, not instead of them. 25 courses in six phases, picked from all 132 in the DeepLearning.AI catalog. Each card shows how solid the mapping is: outline means a lesson title in the course outline names the subject; summary means only the course's “What you'll learn” does.
7Coverage map: AI Engineering from Scratch and Stanford
Each atlas stage against the two main routes. The DeepLearning.AI route is left out because it is a technical add-on, not a full path. A stage is covered when the source names the stage's core concepts, and partly when some are missing.