Learning roadmap¶
English · Русский · Map of the atlas
Choose a route by the work you need to demonstrate. The Atlas still has two tracks: AI Engineering and AI Leadership. Engineering now offers two focused routes; Leadership offers one route with four responsibility branches. The existing eight-week learning path remains the broad baseline. These routes add choices, not three more mandatory courses.
The focused routes assume experienced developers or product, engineering-management and programme practitioners. If the entry checks expose missing fundamentals, remediate first and retry. All durations below are editorial estimates, not validated preparation times. A scheduled week does not count as completed practice when its gate fails.
Choose your route¶
| Your immediate need | Route and scope | Estimated workload | Evidence you should produce |
|---|---|---|---|
| Broad AI foundations or uncertain gaps | Existing learning path, Engineering or Leadership | 8 weeks × 6–8 hours, plus remediation | Explanations, small artifacts and delayed attempts across shared P0 themes |
| Read, change and defend code with permitted AI | AI-assisted coding, within Engineering | 4 weeks × 6–8 hours | Bug diagnosis, context brief, algorithm counterexamples, PR review checklist and design defence |
| Explain a bounded service around a model | Production AI engineering, within Engineering | 8 weeks × 6–8 hours | Baseline choice, response contract, proposed evaluation, retrieval/tool failure analysis and release checklist |
| Lead AI product, teams, portfolios or delivery | AI Leadership, within Leadership | 8 weeks × 6–8 hours for one emphasis | Decision briefs, arithmetic, ownership, release criteria and challenged recommendations |
If you already have recent evidence for an overlapping output, retry it without notes under a changed constraint and use the same gate before substituting it. Reading the baseline does not replace a defensible explanation or decision. The current focused Engineering routes require no Python implementation or lab submission. If you take routes sequentially, add their time budgets and remove overlap only on demonstrated evidence. Training a model, advanced inference work and a full algorithms course are outside these estimates.
Engineering: two different demonstrations¶
AI-assisted coding tests whether you remain in control of a development session. Start with repository orientation and a boundary bug; then explain a scoped AI-assisted feature, challenge algorithm assumptions, and finish with a build plan, review and defence. Seven algorithm families serve as a choice-and-counterexample map, not seven large projects. Use the real interview's tool policy. An employer's published permission does not apply to another employer.
Production AI engineering tests whether the service has a clear task, baseline and failure contract. Keep one small public or synthetic case through model/contract selection, evaluation, retrieval, bounded tools, operations, security and a final changed-constraint defence. Use conceptual examples and proposed checks; mark model quality and security behaviour unmeasured. No model account, API key or implemented service is required. “Production” describes the responsibilities practised, not a certification of readiness.
Both routes reuse the existing Engineering questions. Their editorial follow-ups remain outside the published and radar-generated bank; existing questions now have additional answers and primary reading. The Engineering research report explains the source comparison, Premium-access limits and primary technical checks behind the route design.
Leadership: select responsibility, not title¶
The Leadership route has shared work in problem choice, evidence, technical judgement, authority and delivery. Choose one branch for depth. You do not need to complete every branch's cases within eight weeks.
| Branch | Main decisions | Scope to make explicit |
|---|---|---|
| Product | Discovery, AI/no-AI, metrics/experiments, economics, UX and launch/stop | User and buyer, outcome, evidence, prioritisation authority |
| EM / Platform | Architecture, reliability, AI-assisted SDLC, staffing, hiring and coaching | Service ownership, team capacity, direct reports and escalation |
| Head / Director | Portfolio/budget, build/buy, organisation change and executive trade-offs | Resources actually controlled, managers led, allocation and reversal authority |
| TPM / Delivery | Dependencies, uncertain milestones, readiness, re-planning and handover | Decision owners, commitments, unresolved dependencies and stop conditions |
A Director title does not by itself prove people management or budget authority. Technical depth does not substitute for coaching or portfolio decisions. Practise genuine past examples separately from fictional cases; an exercise cannot become a claimed professional achievement. The Leadership audit shows where interview evidence is direct, secondary, adapted or transferred from another role.
A weekly practice loop¶
- Attempt the question or sketch the artifact before reading. State assumptions and what would disprove your approach.
- Read the selected primary material and existing question outline; correct the artifact rather than memorising an answer.
- Obtain a check from tests, a peer or the cited source. An AI critique needs verification; a numerical AI score alone is not evidence.
- Retry without notes after a delay, changing a relevant constraint. Roughly two days and the following week are practical scheduling choices, not optimised scientific intervals.
- Record date, artifact/question anchor, outcome, score, feedback and next retry. Carry unresolved gaps forward visibly.
Use the existing 0–3 rubric: 0 indicates incoherence or a critical error; 1 needs help or misses a valid check; 2 independently explains or implements the mechanism with assumptions, a trade-off and a check; 3 also handles changed constraints, failure cases and limits. A weekly output and delayed attempt each need at least 2, plus the route's concrete acceptance check. Correctness, access, critical release and truthful-experience failures cannot be averaged away.
The learning path links the research supporting retrieval and spacing and states its limits. That research does not validate this interview curriculum, its schedules or its pass threshold. Completion means recorded practice and successful retries for the chosen scope, not a hiring prediction, course credential or deployment approval.
Read the evidence before expanding the scope¶
The coverage audit separates material coverage from interview evidence and learner proficiency. The expansion plan records implemented content, unresolved evidence and deferred implementation. The current expansion adds questions where provenance permits, answers, reading and review checklists. Python exercises, executable labs, project implementation and an AI evaluator are deferred by the maintainer; they are not requirements of these focused routes. The separate project roadmap remains the repository-wide development plan.