Diagram Sources
The project's diagrams need to be screenshot-ready, maintainable, and bindable to interactive state. Prefer carefully authored SVGs and stable ID conventions.
Recommended Workflow
SVG Naming Conventions
Concept, process, and token nodes.
Data flow, control flow, and influence relationships.
Text labels retained in screenshots and exports.
Current steps, failure states, and resolved states.
Reusable Sources
Demo 01Search Tree: A* Reaches The GoalA* uses heuristic cost to choose the frontier toward Goal
astar-goal-routeStable node / arrow IDsnode-start, arrow-start-b
Teaching simplificationA fixed small tree with unit edge costs illustrates expansion order, not the scale of a real search space.
Demo 02Expert System: Rule ConflictThe same facts trigger two incompatible conclusions
rule-conflictStable node / arrow IDsnode-facts, arrow-facts-rules
Teaching simplificationRules, facts, and the conflict are scripted for teaching and do not represent a real medical or business decision.
Demo 03Bayes: Evidence Raises The PosteriorSupporting evidence moves belief from a 30% prior to a 70% posterior
posterior-raisedStable node / arrow IDsnode-prior, arrow-prior-update
Teaching simplificationThe percentages are simplified to explain the direction among prior, evidence, and posterior.
Demo 04Decision Boundary: Nonlinear SeparationA curved boundary fits two groups while avoiding an outlier
nonlinear-boundaryStable node / arrow IDsnode-negative-cluster, arrow-data-boundary
Teaching simplificationThe two-dimensional points and boundary are hand-authored, not produced by training on real data.
Demo 05CNN: An Edge Kernel Produces A Feature MapA 3×3 edge kernel scans the center window
edge-kernel-centerStable node / arrow IDsnode-image-grid, arrow-image-kernel
Teaching simplificationA single-channel 5×5 grid and fixed kernel omit training, bias, and deeper layers.
Demo 06Feedback Learning: Higher Return Changes Later PolicyAfter two fixed episodes, the right path rises from 30% to 60%; a runtime observation does not update weights automatically
updated-policy-boundaryStable node / arrow IDsnode-state, arrow-state-policy
Teaching simplificationProbabilities, actions, and rewards are scripted teaching states. The diagram performs no random sampling, real reinforcement-learning optimization, or online model training.
Demo 07Attention: Direct Token ConnectionsSelecting model reveals weights to directly and context
selected-token-linksStable node / arrow IDsnode-token-model, arrow-model-directly
Teaching simplificationLine thickness uses scripted weights to explain direct information paths, not real model attention.
Demo 08Foundation Model: Training Versus Runtime ContextThe assistant model and current context produce one answer while weights stay fixed
runtime-context-boundaryStable node / arrow IDsnode-corpus, arrow-corpus-pretraining
Teaching simplificationTraining stages and outputs are scripted for teaching; no real training, preference labeling, or model inference runs.
Demo 09LLM System: Compose External Boundaries By TaskRestored task state flows through a controlled tool and result evaluation
verified-action-pathStable node / arrow IDsnode-task, arrow-task-context
Teaching simplificationComponents, state, actions, and checks are scripted; no real model, memory, tool, or evaluator is called.
Demo 10RAG: From Query To Traceable AnswerRetrieved evidence is reranked and enters an answer with citations
grounded-answerStable node / arrow IDsnode-query, arrow-query-embedding
Teaching simplificationThe nodes and answer are scripted; no vector database, reranker, or model is called.
Demo 11Agent: Revising After A Tool FailureA failed observation enters Revise before the tool retry
retry-after-tool-failureStable node / arrow IDsnode-plan, arrow-plan-tool
Teaching simplificationTool results and retry paths are scripted, with no real external action.
Demo 12Safety / Eval: A Failure Enters The Release GateThe RT-017 fix passes while the old version remains blocked
risk-fixed-regressionStable node / arrow IDsnode-red-team, arrow-red-team-guardrail
Teaching simplificationOne indirect prompt-injection case is shown; passing it does not prove that every system risk is gone.
MVP Core Diagrams
The homepage path through rules, learning, large models, and agents.
DiagramAI Evolution TimelineConnects key technical shifts and matching demos across eras.
DiagramAI Technical LineageMaps symbolic AI, statistical learning, neural networks, foundation models, and agents by paradigm.
DiagramLLM System Boundary MapShows the system relationships among models, context, retrieval, tools, memory, and evaluation.
DiagramSearch Tree Strategy DiagramCompares expansion paths and frontiers for BFS, DFS, and A*.
DiagramExpert System Rule DiagramShows if-then rules, exception conditions, and conflict states.
DiagramBayesian Belief Update DiagramExplains uncertainty updates with priors, evidence strength, and posterior bars.
DiagramDecision Boundary ComparisonCompares linear, nonlinear, and overfit boundaries.
DiagramCNN Kernel ScanShows the local window, convolution kernel, and feature-map response.
DiagramFeedback Learning Policy UpdateShows how delayed return updates an illustrative policy at the training boundary while runtime observations stay separate.
DiagramAttention Token RelationshipsCompares direct token connections with RNN chain propagation.
DiagramFoundation Model LifecycleSeparates pretraining, instruction tuning, preference feedback, and runtime context.
DiagramRAG Pipeline FlowShows Query, Embedding, Vector DB, Reranker, Prompt, LLM, and Answer.
DiagramAgent Loop Action CycleShows planning, tool calls, observation, revision, and the final answer.
DiagramSafety / Eval Release Feedback LoopShows red-team cases, guardrails, least privilege, human review, regression evaluation, and the release gate.
Screenshot Ready
Keep spacing, titles, active states, and simplification notes clearly visible. Before exporting or capturing a screenshot, choose a state that communicates one core aha moment instead of crowding the frame with multiple control panels, long explanations, and complex parameters.
