LEARNING PATH · V0.9

A complete AI Engineering path, without forcing you to start at the beginning.

Follow the 10-stage path when you want structure, or enter through a production incident and backfill only the models you actually need.

01 · PATH

Guided Path

Ten stages from model behavior to production architecture. Stages show the full curriculum even when every mental model does not yet have its own page.

Progress is stored only in this browser for now. No account is required.

  1. STAGE 00

    AI Systems Mental Model

    Understand what the model can suggest and what the application must guarantee.

  2. STAGE 01

    Behavior, Prompt & Output Contracts

    Shape behavior while keeping authority, context and runtime responsibilities explicit.

    Mental models: 4

  3. STAGE 02

    Context, Retrieval & RAG

    Build evidence pipelines that stay relevant, authoritative and inspectable.

  4. STAGE 03

    Memory, Knowledge & Source of Truth

    Design memory as a lifecycle and authority problem, not a storage feature.

    Mental models: 5

  5. STAGE 04

    Tools, MCP & Capability Boundaries

    Give Agents capabilities without turning model intent into permission.

    Mental models: 6

  6. STAGE 05

    Agent Loop, State & Long-Running Work

    Control iteration, state, interruption and work that outlives one request.

  7. STAGE 06

    Reliability, Security & Human Control

    Bound retries, side effects, trust and human intervention under production pressure.

  8. STAGE 07

    Evaluation, Observability & Production Economics

    Use traces and evaluation evidence to decide what is safe and worthwhile to ship.

  9. STAGE 08

    Graphs, Delegation & Multi-Agent Systems

    Add orchestration only when decomposition creates more value than coordination cost.

  10. STAGE 09

    Production Architecture & Capstones

    Integrate the layers and own a defensible production release decision.

02 · MAP

Knowledge Map

Mental models are the durable units. Experiences are practice surfaces that can exercise several models at once.

00AI Systems Mental Model4
  • S00-M01Probabilistic behavior vs application guarantees
  • S00-M02Model claim vs runtime fact
  • S00-M03Structured output as an application contract
  • S00-M04Runtime enforcement vs model persuasion
01Behavior, Prompt & Output Contracts4
  • S01-M01Instruction authority and provenance
  • S01-M02Ambiguity, specificity and instruction conflict
  • S01-M03Examples shape behavior; they do not enforce policy
  • S01-M04Prompt vs Context vs Runtime responsibility boundary
02Context, Retrieval & RAG8
  • S02-M01Finite context budget
  • S02-M02Context structure and cache-friendly stability
  • S02-M03Compaction vs critical-information retention
  • S02-M04Evidence granularity and chunking
  • S02-M05Dense, sparse and hybrid retrieval
  • S02-M06Reranking as a second evidence-selection stage
  • S02-M07Freshness, authority and source priority
  • S02-M08Structured and metadata-aware retrieval
03Memory, Knowledge & Source of Truth5
  • S03-M01Working memory vs durable memory
  • S03-M02Memory write/read lifecycle and selection policy
  • S03-M03Memory vs authoritative source-of-truth conflict
  • S03-M04Expiry, invalidation, versioning and forgetting
  • S03-M05User memory vs domain knowledge boundary
04Tools, MCP & Capability Boundaries6
  • S04-M01Tool contract and schema design
  • S04-M02Tool-result validation and explicit error surfaces
  • S04-M03Capability boundary and least privilege
  • S04-M04Reversible vs irreversible actions
  • S04-M05Human approval at the high-risk boundary
  • S04-M06MCP protocol, application state, authorization and capability boundaries
05Agent Loop, State & Long-Running Work6
  • S05-M01Act → observe → verify loop
  • S05-M02Planning vs direct execution
  • S05-M03Bounded autonomy, termination and escalation
  • S05-M04Checkpoint, resumable state and recovery
  • S05-M05Interrupt and cancellation semantics
  • S05-M06Async and long-running task lifecycle
06Reliability, Security & Human Control5
  • S06-M01Timeout ambiguity: missing response is not confirmed failure
  • S06-M02Retry policy and retry amplification
  • S06-M03Idempotency boundary for repeated intent
  • S06-M04Compensation and recovery after side effects
  • S06-M05Trust boundary and defense in depth for untrusted context
07Evaluation, Observability & Production Economics7
  • S07-M01Traceability as causal execution history
  • S07-M02Observability as a diagnosis interface
  • S07-M03Evaluation environment and verifier design
  • S07-M04Outcome vs trajectory evaluation
  • S07-M05Dataset slices, regression and aggregate-improvement traps
  • S07-M06Confidence, variance and sample-size humility
  • S07-M07Cost, latency, quality and release vetoes as one decision
08Graphs, Delegation & Multi-Agent Systems4
  • S08-M01Loop vs Graph responsibility boundary
  • S08-M02Sequential, conditional and parallel topology trade-offs
  • S08-M03Delegation contract and shared vs isolated state
  • S08-M04Independent verification, coordination cost and correlated failure
09Production Architecture & Capstones3
  • S09-M01Cross-layer architecture decomposition
  • S09-M02Bounded rollout, fallback and graceful degradation
  • S09-M03Evidence synthesis into SHIP / BLOCK / INCONCLUSIVE