CAREER
Quality-AI logoQuality-AI Hiscox SoleraSmart Drive Accion Labs KPMG Dell Technologies Wells Fargo Virtusa logoVirtusa Qatar Airways logo Citi Bank logo JD logoJD Sports Conduent Bank of America MetLife
THE COMPLETE LEARNING BLUEPRINT / 2027+ MASTER EDITION

From software engineering, technical leadership, cloud, data, machine learning and AI roles to Enterprise AI Solution Architect.

A complete learning path from Python and machine learning to multi-agent systems, durable workflows and production AI. Build with open models, measure quality, enforce budgets, deploy securely, and defend every architectural decision.

Explore 231 chapters →View learning path ↗
19Learning parts
60Core modules
231Detailed chapters
18Portfolio systems
00 / NORTH STAR

What you’re becoming

An AI Solution Architect translates a business objective into an effective, secure, scalable and cost-aware AI system. This curriculum connects programming, mathematical foundations, machine learning, advanced AI, cloud platforms and production architecture in a dependency-based sequence. It is a mastery-oriented roadmap, not an official certification syllabus.

How to study: For every important concept, explain the intuition → derive the mathematics → implement it from scratch → use modern frameworks → test, debug and deploy → defend architectural trade-offs. Follow the numbered parts: foundations → machine learning → advanced AI → retrieval → agents → durable workflows → operations → cloud delivery → solution architecture. Build a project at each stage and use its evidence gate before moving on.
2027+ MASTER PATH / READ → BUILD → PROVE

A sequence you can actually follow.

Move through six delivery stages. Every stage connects numbered curriculum parts to a project and a concrete readiness check. Learn one working stack deeply; comparison tools are optional reading.

2027 & beyond · built on verifiable foundationsPrepared 20 October 2026. Model and product examples are verified baselines, not predictions of future releases. Before each major project, review official documentation, licenses, security advisories and model support; pin the versions you test.
CORE IMPLEMENTATIONLangGraph+Temporal+pgvector+LangSmith+OpenTelemetry→One cloud, operated well
LEARN BY SEEING / INTERACTIVE CONCEPT LAB

Make the mathematics move.

Interactive visual lessons connect formulas to algorithms, neural architectures, retrieval, agent frameworks, MLOps and cloud AI platforms. Computed examples use fixed synthetic data; architecture and product workflows are labeled schematics. The hero rotates through the complete visual lesson library.

Swipe or scroll sideways to explore the full diagram.

CORE FORMULA

Technical reference ↗
HOW IT WORKS
    Study this concept ↗

    Animations play once per lesson. Pause at any time.

    01 / THE COMPLETE SYLLABUS

    A clear sequence. From foundations to architecture.

    All 231 chapter titles and detailed topics are included below. Open a module to inspect its syllabus. Search across chapter names, concepts and technical terminology. Mark individual topics as you learn, or complete an entire chapter at once. Your progress saves automatically in this browser. Export a backup to move topic and project progress to another browser or device.

    Showing all 231 chapters

    Saved on this browser · Export regularly to keep a portable backup.

    OPEN-MODEL LAB / COLAB TO A CONTROLLED SERVICE

    Choose a model for the task.

    Twelve model choices across text, reasoning, coding, vision, image and video generation, speech and retrieval. Start with the three text baselines for P07, then select a specialist for your project. These are practical learning choices, not a universal ranking or a forecast of the best models in 2027. Each repository is linked below; read its license and current loading instructions.

    12 model choices

    Coverage map · what to study before choosing a model

    This is a broad enterprise AI pathway. No finite syllabus covers every concept or library; use these checkpoints to locate the core subjects and the task-specific extensions.

    Foundations

    Python, arrays, dataframes, linear algebra, calculus, optimization, probability, statistics

    Python & the AI Library Ecosystem ↗Mathematics for Machine Learning ↗Probability & Statistics ↗
    Learning algorithms

    Regression, classification, clustering, ensembles, validation, gradients and neural networks

    Machine Learning: Concepts & Algorithms ↗Deep Learning & Neural Networks ↗
    Language & multimodal AI

    Tokenization, transformers, attention, adaptation, diffusion, speech, vision and video

    Natural Language Processing & Language AI ↗Computer Vision & Multimodal AI ↗Transformers, LLMs & Generative AI ↗Modern AI Practice & Emerging Patterns ↗
    Retrieval & agent systems

    Embeddings, reranking, vector indexes, typed tools, state graphs, coordination and durable execution

    Vector Databases & Retrieval Engineering ↗AI Agents & the LangChain Stack ↗Multi-Agent Systems & Orchestration ↗Durable Agentic Workflows with Temporal ↗

    For each selected stack, study the linked tool registry, complete the project evidence gate, and record package versions and model revisions. The multimodal chapter now explicitly includes speech metrics, sampling, diffusion pipelines and memory planning.

    Memory planning: FP16 weight estimates are parameter count × 2 bytes for a single model, not the memory for a complete multimodal pipeline. KV cache, activations, runtime overhead and fine-tuning require additional memory. Start with batch size 1, a 2,048-token input limit and 256 output tokens; reduce these if needed. GPU type and availability vary in Colab. A 16 GB GPU is a useful lab target, not a guarantee of fit.
    01 · ESTABLISH

    Inspect & pin

    Record GPU memory, package versions, model revision, tokenizer, license, chat template and dataset split. Keep credentials out of shared notebooks.

    02 · COMPARE

    Run the same test set

    Measure schema validity, task success, input/output tokens, wall time and peak memory on 30 fixed tasks. Load one model at a time.

    03 · ORCHESTRATE

    One model, three roles

    Reuse one loaded model for researcher, analyst and reviewer calls. Start sequentially; distinguish logical roles from parallel model replicas.

    04 · PROMOTE

    Leave the notebook

    Move validated code into an API and workers. Run Temporal and persistent databases outside the disposable Colab runtime. Re-test the deployed model.

    For the retrieval baseline, use sentence-transformers/all-MiniLM-L6-v2 ↗ as a separate 384-dimensional embedding model. It is not one of the three chat models. Benchmark domain relevance and truncation behavior before relying on it.

    Open the model benchmark project →
    REFERENCE DESIGN / BOUNDED AUTONOMY, DURABLE EXECUTION

    From an agent loop to an operated system.

    This reference architecture assigns one owner to each responsibility. Use fewer components for early projects, then add durability and operational controls when the use case requires them.

    01

    Request & policy

    FastAPI · identity · tenant scope · budget reservation

    02

    Durable workflow

    Temporal · timers · approvals · retries · recovery

    03

    Agent decisions

    LangGraph · typed state · specialist roles · stop policy

    04

    Models & tools

    Shared inference · permission-aware retrieval · validated actions

    05

    Evidence & usage

    LangSmith · OpenTelemetry · usage ledger · dashboards

    LangChain / LangGraph

    Model and tool integration, agent state, routing, checkpoints and human interrupts. Keep deterministic policies outside model prompts.

    Runtime documentation ↗

    Temporal

    Long-running business coordination. Put model/network calls in Activities, design idempotent external effects, and choose one retry owner at every boundary.

    Workflow documentation ↗

    LangSmith + OpenTelemetry

    Agent-level evaluation and end-to-end operational evidence. Correlate request, workflow and agent IDs, and redact sensitive content before export.

    Tracing documentation ↗
    Autonomy is a permission level. Begin with read-only tools. Add scoped writes only after schema checks, authorization, idempotency and approval paths pass tests. More agents do not automatically improve quality; compare against one agent and a deterministic workflow.

    Token tracking is evidence. Budget enforcement is control.

    ReserveEstimate input + maximum output; atomically reserve against the shared tenant budget.
    ExecuteApply tool, step, time and output-token limits; propagate cancellation to workers.
    ReconcileUse provider-reported usage when available; label estimates and de-duplicate events.
    ReportCompare ledger totals with invoices and infrastructure spend; alert on anomalies.

    API cost model: (uncached input tokens × input rate + cached tokens × cached rate + output tokens × output rate) / 1,000,000 when rates are per million tokens. Include retries, tools, storage and GPU time separately. Use dated price snapshots; this page does not assume future provider prices. See cost tracking and gateway budgets.

    Choose your retrieval layer deliberately.

    FIRST BUILD

    PostgreSQL + pgvector

    Keep application records, metadata and vector queries close together. Benchmark filters, indexes and tenant boundaries.

    Study & official reference →
    SECOND IMPLEMENTATION

    Qdrant

    Compare a dedicated vector service on the same corpus and query set. Include snapshots, ingestion and operating cost.

    Study & official reference →
    COMPARISON ONLY

    Managed & distributed options

    Read about Pinecone, Weaviate and Milvus when scale or operating constraints justify them. FAISS remains a local index baseline.

    Compare the alternatives →
    CLOUD AI / FROM SERVICES TO SYSTEMS

    Three clouds. One architecture mindset.

    Study the dedicated Azure, AWS and Google Cloud tracks after the modeling foundations. Provider names and catalogs evolve; these links point to official service documentation reviewed on 20 October 2026.

    Open the dedicated Azure services + .NET learning path → Open the dedicated AWS services + .NET learning path →

    Cloud study includes identity, private networking, data protection, deployment, evaluations, monitoring, resilience and cost control—not just service names. Google Cloud uses Gemini Enterprise Agent Platform (formerly Vertex AI); retain legacy names when reading older examples.

    02 / CONTINUOUS PRACTICE

    Your cross-disciplinary toolbelt.

    Learn underlying patterns instead of memorizing tools. Become deeply competent with one cloud provider and develop working fluency in alternatives.

    A focused delivery stack. Browse all tools, roles, symbols and official references →

    Skills across the stack

    AreaEssential technologies and skills
    03 / ARCHITECTURE MINDSET

    Design the whole system.

    Example: an enterprise AI assistant answers employee questions using internal documentation. A model engineer improves embeddings, prompts and retrieval. A solution architect owns the data flow, controls, integration, reliability, operational model and measurable business outcome.

    ILLUSTRATIVE ENTERPRISE RAG ARCHITECTURE
    Enterprise usersPortal, chat, Teams
    Identity & APISSO, access policy, rate limits
    OrchestrationPrompts, tools, guards
    Retrieval + LLMAuthorized search, answers
    Data sourcesDocuments, databases
    CROSS-CUTTING: Security · Observability · Evaluations · Cost · Reliability · Audit · Human escalation

    Conceptual request-path view, not a complete deployment topology. Production designs need separate ingestion/indexing pipelines, document-level authorization at retrieval time, resilience and governance controls.

    Questions an architect must answer

    06 / CAREER CALIBRATION

    Depth where it matters most.

    Deep Learning fundamentalsDeep understanding
    NLP & LLM EngineeringDeep understanding
    Computer VisionStrong practical competence
    Cloud, MLOps & distributed systemsDeep practical competence
    Security, governance & architectureExpert-level target
    The goal is not simply knowing 231 chapters. It is the ability to take an enterprise AI use case from requirements through solution design, proof of concept, production deployment, monitoring and continuous improvement—and explain exactly why every architectural component is there.
    07 / CANONICAL REFERENCES

    Learn from official sources.

    Documentation and primary learning material to use alongside practical labs and projects. External links require internet access.

    BUILD PORTFOLIO / 18 PROJECTS WITH EVIDENCE

    From your first model to autonomous workflows.

    Each project identifies what to study, which tools to use, what to deliver and how to prove it works. These are build briefs—not claims that the systems have already been implemented. Follow the prerequisite projects and keep your evidence alongside your code.

    Project completion saves in this browser and is included in progress exports. Mark a project complete after its evidence gate passes.

    FINAL REFERENCE / TOOLS, SYMBOLS & OFFICIAL DOCUMENTATION

    Know what each tool is here to do.

    The selected implementation stack, its lab tools and explicit alternatives—all in one place. Core means used on the main learning path; lab means a bounded experiment; comparison means optional reading; choose one means a deployment decision. Do not install everything at once.

    CoreLabComparisonChoose one cloud

    Bundled project marks use Devicon and Simple Icons where available. A labeled category symbol is used for entries without a bundled brand mark. Every entry links to its official documentation. Names and trademarks belong to their respective owners. Cloud service-level references are available through each provider’s catalog.

    Keep this curriculum current for 2027 and beyond
    • Before a project: review model cards, licenses, runtime compatibility and security advisories.
    • Pin package versions, model revisions, schemas and evaluation datasets in a project manifest.
    • Re-run quality, latency, failure, privacy and budget tests before promoting any new model or framework version.
    • Review cloud services, deployment limits and provider pricing using official sources; keep a dated decision record.
    • Remove a tool when it no longer solves a measured requirement. Keep alternatives as reading until a project needs them.