Practical AI · Personal learning roadmap
Start where you are.
Learn what matters for your work.
Choose your experience and profession. Get a focused sequence of skills, practice tasks, and evidence that you are ready to move on.
No sign-up. A curated, rule-based guide—not an AI assessment or a certification. Your selections are processed in this page, not sent to an AI service.
Interactive controls need JavaScript. You can still use the starting-point and profession guides below.
Your learning path
Protect sensitive data, verify sources, test failures, and retain human approval for consequential actions. Experience does not exempt you from these checks.
Printing uses your browser. This page does not save your progress.
Browse without the interactive finder
Choose your starting point
I’m new to AI: Start with foundations and one safe, useful task. Work through the four foundation steps, then your profession project.
I use AI occasionally: Make your results more reliable and repeatable. Check your foundations, improve your task brief, then learn grounding and evaluation before your profession project.
I use AI confidently: Test your skills and design a bounded workflow. Pass the foundation checks, then focus on evidence, evaluations, and your profession project.
I already build AI solutions: Deepen evaluation, security, and operations. Coding experience assumed. Verify prerequisites, then focus on your profession’s advanced work and a tested prototype.
Foundation sequence
- Understand the tool before trusting it. Distinguish prediction, generation, retrieval, and actions. Learn why fluent answers can be wrong. Try: Ask an approved assistant to summarize a short public document. Mark each claim supported, unsupported, or uncertain. Ready when: You can explain two limitations and trace every accepted claim to its source.
- Write a task brief, not a magic prompt. Specify the goal, audience, trusted context, constraints, examples, and output format. Try: Improve the same task using a written brief. Compare both answers against five criteria you set before running it. Ready when: You have a reusable brief and can explain which changes improved the result.
- Ground answers and handle missing evidence. Select current, authoritative sources; verify citations; distinguish manual grounding from an implemented RAG pipeline. Try: Use two conflicting synthetic policies plus a question neither answers. Record how your assistant handles each case. Ready when: It uses the current policy, identifies the conflict, and does not invent a missing rule.
- Evaluate a repeatable workflow. Create representative tests, record expected results, inspect failures, and retain human approval for impactful actions. Try: Build ten cases: normal requests, missing evidence, ambiguity, stale sources, hostile instructions, and unauthorized actions. Ready when: You record actual outcomes and reruns—not just a completion claim or a confidence score.
Find your profession
Business, product & operations
Learn: Workflow mapping, document/data verification, process ownership, and pilot economics.
First project: Create a synthetic policy-question assistant that drafts answers but cannot approve or execute requests.
Evidence: Compare five manual tasks with five assisted tasks, including checking and rework. Record errors and time without claiming causal proof.
Builder extension: Instrument a bounded workflow with provenance, versioning, time/cost limits, escalation, and rollback.
Defer: Model training, a front-end framework, and autonomous multi-agent systems.
Software development & architecture
Learn: One language, HTTP/JSON/APIs, Git, testing, structured outputs, retrieval, tool contracts, and authorization.
First project: Build a small source-grounded assistant using public documents and a read-only tool. Keep secrets server-side.
Evidence: Test schema errors, absent evidence, injected instructions, denied access, and timeouts; inspect actual state changes.
Builder extension: Add an evaluation harness, trace-based diagnosis, scoped tool permissions, idempotent writes, budgets, and rollback.
Defer: Fine-tuning and elaborate agents until a simpler baseline fails on measured tests.
QA, testing & quality engineering
Learn: Expected outcomes, representative test sets, regression, repeated trials, human-review calibration, and failure diagnosis.
First project: Create an evaluation pack for a synthetic policy assistant, with a separate answer key and unseen variants.
Evidence: Distinguish retrieval, generation, tool, and permission failures; rerun unstable cases and report untested controls honestly.
Builder extension: Automate outcome checks, version datasets and graders, calibrate human/automated judgements, and gate releases on critical failures.
Defer: A single model-as-judge score as the sole release gate.
Data & analytics
Learn: Data quality, units, missingness, SQL/statistics foundations, chart interpretation, provenance, and uncertainty.
First project: Analyze a synthetic table with known totals, missing values, and one misleading pattern. Check AI claims against calculations.
Evidence: Reconcile totals and units; label uncertainty and distinguish association from causation.
Builder extension: Build reproducible analysis and evaluation pipelines, access-controlled retrieval, dataset versioning, and drift checks.
Defer: Training a large model before you have a clean dataset and a defensible baseline.
Education & training
Learn: Learning objectives, source checking, accessible materials, formative assessment, and academic-integrity boundaries.
First project: Draft a lesson and assessment using public material, then check every factual claim and answer key.
Evidence: A colleague can teach from the guide; an unseen exercise tests understanding rather than copying.
Builder extension: Evaluate a source-grounded teaching assistant with misconception tests, accessibility checks, retention rules, and teacher review.
Defer: Automated high-stakes grading or entering identifiable student records into unapproved tools.
Healthcare & life sciences
Learn: Evidence currency, privacy, domain validation, bias, human oversight, and safe administrative use.
First project: Draft a public-source administrative FAQ using synthetic scenarios—not patient records or diagnosis requests.
Evidence: Verify every answer and limitation with a qualified reviewer; refuse unsupported clinical recommendations.
Builder extension: In a governed sandbox, test source freshness, access isolation, auditability, escalation, and change control with domain/privacy/security reviewers.
Defer: Autonomous diagnosis, treatment decisions, or real patient data in unapproved systems.
Leadership & management
Learn: Use-case selection, build/buy decisions, accountability, adoption, vendor/data risk, and total cost per accepted task.
First project: Write a bounded pilot charter with a non-AI baseline, owner, users, success measures, and stop/go criteria.
Evidence: Include review effort, failure costs, residual risks, and evidence needed before expansion—not just vendor ROI claims.
Builder extension: Design portfolio governance, evaluation gates, incident ownership, model/vendor change reviews, and a measured adoption process.
Defer: Choosing an agent or platform before identifying the problem and decision owner.
Other profession / exploring
Learn: Task framing, trustworthy research, document/data checks, safe reuse, and human decision boundaries.
First project: Choose one low-risk weekly task. Use only public or synthetic inputs, and keep the original manual process as a baseline.
Evidence: Show the evidence behind the output, one failure you caught, and whether review effort offsets the benefit.
Builder extension: Prototype a bounded workflow only if you can specify permissions, expected results, failure handling, and an accountable owner.
Defer: Specialist infrastructure or autonomous actions without a demonstrated need.
Resources and how recommendations work
The guide combines a shared foundation with a profession-specific project. Your selected level changes the sequence; it does not verify your competence. Use the readiness checks before skipping a topic. Time to proficiency varies by prior knowledge and project complexity.
- AI Concepts — Sakti Bagchi
- Context engineering — Anthropic
- RAG Explained — Sakti Bagchi
- Building effective agents — Anthropic
- Agent evaluation — Anthropic
- AI Agent Security — OWASP
- Responsible coding-agent use — GitHub
These are starting resources, not a comprehensive or ranked course list. Provider guides are implementation perspectives, not independent proof of learning effectiveness. Recommendations adapt the practical AI curriculum reviewed on 16 November 2026; recheck living documentation before building.