82 templates
AI & Pipelines diagram templates
RAG, training pipelines, agentic workflows, model orchestration.
RAG Pipeline
Documents → embed → vector DB; query → retrieve → LLM → answer.
WeaveML training pipeline — feature engineering and split
Raw data → cleaning → features → split → train → evaluate → deploy.
WeaveLLM agent loop
User → planner → tools → observation → response.
WeaveLive event stream (animated)
Producers → Kafka stream → consumers; the edges along the hot path FLOW like the landing hero.
WeaveAI orchestrator (live)
User → router → tool agents → response; pulses light the active branch.
WeaveRAG pipeline
Ingestion path + query path meeting at retrieval.
D2Learning path: ML engineering
Roadmap of topics with time estimates.
Markmap (Markdown)AI product strategy
User-facing AI app on top of frontier models, embeddings, vector DBs.
Wardley MapsVenn — AI × Data × Product
Where AI features actually require all three disciplines.
Venn / EulerVenn — RAG vs. fine-tuning vs. prompt
Where these three customisation strategies actually overlap.
Venn / EulerVenn — data-science skill triangle
Drew Conway's classic data-science Venn updated.
Venn / EulerPERT — ML pipeline build
Standing up a training pipeline from scratch.
PERT / CPM networkSwimlane — AI feedback loop
User / Frontend / Backend / Eval team feedback cycle.
Swimlane processStructurizr — AI agent platform
Agent execution platform with orchestrator, tool router and memory.
Structurizr DSLWardley map — LLM-native app
Where each LLM-app layer sits on the evolution axis.
Wardley MapsD2 — Gen-AI application
RAG-style architecture with retriever and reranker.
D2Argument map — will AI cause mass unemployment?
Pro / con / rebuttal structure on AI labour displacement.
Argument mapRAG + LLM Stack
Documents embedded by a neural encoder into a vector database, retrieved as top-k context for a transformer LLM that answers the query.
WeaveCNN Image Classifier
A convolutional pipeline — an image through stacked conv layers and pooling, flattened into a dense network and a softmax over classes.
WeaveTransformer Architecture
Tokens are embedded, passed through a stack of transformer blocks, then projected and softmaxed into the next-token distribution.
WeaveMLOps Pipeline
Data flows into a feature store, a training job writes to a model registry, serving pulls the model, and monitoring closes the retrain loop.
WeaveGAN (adversarial)
A generator turns noise into fakes while a discriminator learns real from fake; the loss back-propagates to both networks.
WeaveQuadrant — LLM Cost vs Quality
Inference cost × output quality for model routing decisions.
Quadrant MatrixQuadrant — ML Use-Case Selection
Business value × technical feasibility for an AI/ML roadmap.
Quadrant MatrixQuadrant — Data Quality
Completeness × accuracy to triage data sources before modelling.
Quadrant MatrixQuadrant — Prompt Latency vs Quality
Response latency × answer quality when tuning prompts and pipelines.
Quadrant MatrixQuadrant — Feature Importance vs Freshness
Predictive importance × update frequency for feature-store design.
Quadrant MatrixCFD — ML Experiment Queue
Twelve days of a shared GPU cluster's training queue, contrasting a 4.5-day committed cycle time with a submission backlog that grows tenfold, and breaching the concurrent-training limit.
Flow metrics (CFD)DFD — RAG Support Assistant
Retrieval-augmented support assistant from knowledge-base ingest to grounded answer, isolating the third-party embedding and completion APIs so the 8 flows leaving the trusted VPC — prompt text included — are counted as attack surface.
Data flow diagramDSM — Retail Recommender Platform
Component DSM of a retailer's recommendation stack showing why an online learning system cannot be fully sequenced: serving, metrics, drift, retraining and the model registry close a loop that absorbs 10 of the 14 components into a single iteration block, leaving only event collection, identity resolution, the feature store and experiment assignment cleanly upstream.
Design structure matrixDSM — Pre-Training Parameter Coupling
Parameter DSM of the fourteen design choices behind a foundation-model pre-training run, read in inputs-in-columns convention and taken in config-file order, computing that 12 of the 31 couplings run backwards there and resolving the parameters into three blocks: data mixture with vocabulary, context and evals; hidden width with layer depth; and the sharding/optimiser/batch/checkpointing memory loop.
Design structure matrixLean Canvas — Clinical Coding AI Copilot
Lean canvas for an AI copilot that drafts clinical coding for NHS trusts, pricing 310,000 episodes a month against GPU inference and safety-case costs, with coder acceptance rate and hallucinated-code rate as the metrics that decide the product.
Business model canvasStory Map — Adaptive Learning Platform
A first-year plan for an adaptive study platform using ordinal band tokens, where the mastery model and retrieval-grounded tutor land only in band three and the pilot cohort ships with no way to ask for help at all.
User story mapStory Map — Telehealth Urgent Care Triage
Remote-first urgent care where the first band is clinician-led and complete across all four activities — a real walking skeleton — while model-suggested triage, red-flag escalation and connected devices arrive later, two of them still unestimated.
User story mapEvent Storm — Content Publishing with Model Moderation
Newsroom draft through classifier-based moderation to distribution, where the human moderator is declared but sits on no command — the review loop nobody modelled.
Event stormingEvent Storm — Agentic Ticket Refunds
Support desk, an LLM refund assistant and finance approval, with the assistant's escalation event left uncommanded and a policy-retrieval hotspot on stale guidance.
Event stormingRisk Matrix — AI Customer Assistant Risk Register (NIST AI RMF)
An AI risk register for a fictional utility's LLM billing assistant, mapped to NIST AI RMF and EU AI Act transparency duties: across nine failure modes from wrong tariff answers to prompt injection, controls cut exposure from 99 to 45 (−55%) and take the Critical prompt-injection risk to Medium, leaving missed vulnerable-customer handoffs as the one High risk above appetite. Illustrative register.
Risk matrixDecision Matrix — Document AI Model Selection
Four invoice-extraction approaches benchmarked on 2,000 held-out invoices, with accuracy, latency and cost per 1,000 pages measured and a 95% field-accuracy must-have that rules out the current OCR-and-rules pipeline. The fine-tuned open model ranks first at 4.11 of 5, just 0.09 ahead of the small hosted model — a fragile lead that flips if ops effort's weight rises from 10% to 12.5%. Illustrative values.
Decision matrix (weighted / Pugh)Roadmap — LLM Support Assistant From Pilot to General Availability
A month-by-month roadmap for taking a fictional retailer's customer-support LLM assistant from an internal pilot to general availability, with evaluation, retrieval, safety and cost work in their own lanes: 16 items (4 done, 4 in progress, 7 planned, 1 at risk) and 3 milestones, status date 19 Apr 2027. All 21 dependencies finish before their dependents start — the customer beta waits for red-teaming and the DPIA, and GA waits for online evaluation — and the longest chain, six steps, runs from the golden set through CI evals, model routing and the semantic cache to GA and the voice pilot. Illustrative roadmap.
RoadmapRAG Pipeline (Retrieval-Augmented Generation)
Indexing and query paths for a retrieval-augmented LLM app.
MermaidNeural Network Architecture (Graphviz)
A simple feed-forward network's layers.
Graphviz (DOT)Agentic AI Workflow (D2)
A planning agent that calls tools in a loop until done.
D2Machine Learning Pipeline
From raw data to a monitored, deployed model.
MermaidML Model Serving (D2)
Online inference with a feature store and monitoring.
D2Fishbone — Fraud Model Precision Drop
Root-cause analysis of a card fraud model whose precision fell six points in a quarter, grouped across data, features, labels, serving, monitoring and the fraud population itself.
Fishbone (Ishikawa)RAG Evaluation Harness
Splits retrieval and generation metrics against a frozen 420-question golden set so a regression is attributed to the right stage.
MermaidCredit Scoring Model Lifecycle
A regulated model's states from development through independent validation, shadow running and PSI-triggered recalibration or redevelopment.
MermaidMotor Claim Triage Agent (Tool Routing)
A tool-calling agent that pulls policy and fraud signals on first notification of loss, then escalates to a human adjuster above threshold.
MermaidGuardrailed LLM Request Path (PlantUML)
The full inbound and outbound guardrail chain around a retrieval-augmented answer, including citation verification and audit logging.
PlantUMLFeature Store Architecture (D2)
Streaming and batch pipelines feeding an online and offline feature store, with a drift monitor closing the retraining loop.
D2Nightly Training DAG (Graphviz)
A demand forecasting retraining DAG with a data-quality suite, time-series splits and a WAPE quality gate that pages the owner on failure.
Graphviz (DOT)Serving Latency Percentiles (Plotly)
Time-to-first-token at p50, p95 and p99 across five serving configurations against a two-second tail budget.
PlotlyToken Spend by Workload (Vega-Lite)
Monthly inference spend stacked by workload, showing batch summarisation overtaking interactive chat once it moved to a larger model.
Vega-LiteModel Accuracy versus Cost (Vega-Lite)
Candidate serving configurations plotted on a log cost axis against exact-match accuracy, with the routed setup sitting on the frontier.
Vega-LiteModel Delivery Board (Kanban)
A risk and pricing squad's board where model risk review and a sixty-day shadow run are first-class columns, not afterthoughts.
Kanban boardAI Use-Case Portfolio - Feasibility vs Value (Quadrant)
Eight insurer AI use cases scored on today's feasibility against business value, with autonomous settlement parked as high value but not yet buildable.
Quadrant MatrixRecommender Exposure and Purchase (Causal DAG)
Why naive click-through attribution overstates a recommender: item popularity drives both what gets shown and what gets bought.
Causal DAGFlame Graph — PyTorch Training Loop, Main Process (py-spy)
Where the Python main process of a ResNet-50 training run spends its time (4,760 samples): 45% blocked waiting for the DataLoader queue, so the GPU is starved by JPEG decoding in too few workers, and a further 15.1% stalled in a per-step loss.item() that forces a CUDA synchronise. Forward, backward and the optimiser together are only 33.5%. Illustrative values.
Flame GraphML — ablation study
ML evaluation vertical — models, datasets, evals, ablations. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptML — retrieval-augmented LLM ablation study
Component ablation for a retrieval-augmented LLM: each row removes or swaps one part of the pipeline and reports the change in exact-match accuracy, ranked by size, with protocol, search budget and compute stated. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptRetrieval pipeline ablation
Compare removals from an evidence retrieval pipeline, from the permission prefilter to the citation verifier, against its fixed baseline; illustrative deltas are not measured performance or causal evidence. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptDemand forecasting ablation
Compare removals from a demand forecast, from seasonality inputs to regional pooling, against its fixed baseline; illustrative deltas are not measured performance or causal evidence. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptVisual inspection ablation
Compare removals from a surface defect detector, from illumination normalisation to uncertainty routing, against its fixed baseline; illustrative deltas are not measured performance or causal evidence. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptDocument triage ablation
Compare removals from a document classifier, from section structure to abstention, against its fixed baseline; illustrative deltas are not measured performance or causal evidence. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptSpeech transcription ablation
Compare removals from a speech recogniser, from noise augmentation to speaker adaptation, against its fixed baseline; illustrative deltas are not measured performance or causal evidence. Typed FlowScript for machine-learning evaluation. Keywords: ablation, benchmark, evaluation, model comparison.
FlowScriptBowtie — unsafe or non-compliant LLM response to a customer
Barrier-based bowtie for a customer-facing LLM assistant emitting unsafe or non-compliant output, with preventive barriers against injection, jailbreak, stale retrieval and ungrounded generation and recovery barriers for regulatory breach and PII disclosure, used in an AI safety case or Consumer Duty review.
Bowtie riskConfusion matrix — intent classifier held-out set
Row-normalised 6x6 confusion matrix with raw counts for a customer-support intent classifier on a 4,000-utterance held-out set, the figure an ML engineer files with every classifier release to show per-class recall and the dominant confusions.
Vega-LiteReliability diagram — verifier calibration with ECE and MCE
Calibration curve of predicted probability against observed rate in ten bins with confidence intervals and a bin-count histogram, reporting expected and maximum calibration error, used to set an abstention threshold and to sign off a probabilistic model.
Vega-LiteBias audit — selection-rate impact ratios (four-fifths rule)
Selection rate and impact ratio by sex and race/ethnicity for an automated employment decision tool against the 0.80 four-fifths threshold, the disparate-impact table required by NYC Local Law 144 and the EEOC Uniform Guidelines.
Vega-LiteGrounded answer with evidence review
Traces permissions and claim-level support through a retrieval workflow rather than implying retrieval guarantees truth. Original worked design example; adapt responsibilities, thresholds and assumptions to your system.
MermaidModel validation and monitored rollout
Separates offline evaluation, shadow traffic, canary rollout and later drift response. Original worked design example; adapt responsibilities, thresholds and assumptions to your system.
MermaidModel registry and reproducible evaluation
Connects immutable model artifacts to evaluation configurations, approvals and observed deployment windows. Original worked design example; adapt responsibilities, thresholds and assumptions to your system.
MermaidAgent tool authorization and result verification
Limits delegated tools and ties consequential approval to the exact action payload. Original worked design example; adapt responsibilities, thresholds and assumptions to your system.
Graphviz (DOT)Forecast production with leakage and drift checks
Uses time-aware evaluation and monitors interval coverage as well as point accuracy. Original worked design example; adapt responsibilities, thresholds and assumptions to your system.
MermaidAI release gate with quality, safety and data sign-off
Join independently documented evaluation, misuse testing and data provenance review before promoting a model release. Editable collaboration with role lanes, decisions and independent parallel checks. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0Data contract review with schema repair loop
Review schema compatibility, ownership and quality expectations; allow a producer to correct a version or withdraw it. Editable collaboration with role lanes, decisions and a correction and resubmission loop. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0Model card review with evidence correction
Require performance claims, known limitations and intended-use statements to match traceable evaluation evidence before publication. Editable collaboration with role lanes, decisions and a correction and resubmission loop. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0AI provider incident with evidence response deadline
Request provider incident evidence, verify impact containment and route an overdue or incomplete response to a controlled model fallback. Editable collaboration with role lanes, decisions and an interrupting response timer. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0Independent model evaluation acceptance window
Seek independent evaluation acceptance before a deadline; validate evaluator independence before opening a controlled testing workspace. Editable collaboration with role lanes, decisions and a response-versus-expiry event race. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0Cloud environment provisioning with cleanup on failed checks
Provision a scoped environment, validate security controls, and remove created resources if acceptance checks fail. Editable collaboration with role lanes, decisions and explicit rollback activities. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0Dataset migration with validation and pointer rollback
Copy data into a candidate store, verify integrity and only then switch the consumer pointer; failed validation leaves the original dataset authoritative. Editable collaboration with role lanes, decisions and explicit rollback activities. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0Partner API onboarding with credential revocation on failure
Provision scoped partner access and validate conformance before activation; failed onboarding revokes credentials and removes provisional routing. Editable collaboration with role lanes, decisions and explicit rollback activities. Illustrative policy and timing; adapt to local requirements.
BPMN 2.0