AI & DATA SCIENCE / DELIVERY METHOD
From use case to useful change
The strongest work starts with a decision and ends with an accountable change in practice.
Find the decision
Listen to users and service owners, map the current flow, locate delays or uncertainty and agree what a useful intervention would change. For an internal consultancy, I also clarify who owns the decision and which benefit is a proposed target rather than an observed result. Procurement approval bottlenecks and uneven Welcome Week survey structures started as workplace problems, not model briefs.
Be explicit about provenance
Use governed real data only when access and purpose permit, and publish only proportionate summaries. Ask about source, owner, quality, migration and permissions in ways that fit the project; a single data-requirements template cannot serve every system. Where they do not, curate synthetic records that resemble the workflow and label the exercise as a demonstrator. Keep concepts with no dataset in a separate category. Avoid treating a synthetic score as proof of real-world impact.
Make information comparable
Document definitions, missingness, transformations and quality checks. In Welcome Week, the annual survey fields had to be harmonised before trends could be read. In software usage analysis, licence types and activity dates needed consistent recency logic.
Start with a baseline
Compare simple rules and interpretable models with more complex methods. For an agent PoC, define functional scenarios, missing-data and no-permission responses, and the boundary between controlled tests and a supported service. Check class-level errors, leakage, stability and whether the model answer matches the actual decision. A high accuracy score alone is not an outcome. An impressive decimal point still needs a useful job.
Put a human action beside the signal
Explain what a pattern suggests, what it cannot establish, who reviews it and what happens next. A licence recommendation needs administrator review; a student experience correlation invites investigation; a people-analytics signal never becomes an automatic appraisal.
Make the change repeatable
Give the team a usable output, feedback route, owner and measure. Reusable activity catalogues, process standards and learning pathways help the method travel across projects. Re-run the analysis when data changes, check whether the intervention helped, and revise the method. The goal is an enduring capability, not a one-off notebook.
READ THE EVIDENCE CORRECTLY
Three data bases, three kinds of claim.
Shows patterns in a defined workplace source. It still needs privacy, context, quality and appropriate review.
Welcome Week · BA feedback · software usageTests workflow and technical feasibility without exposing live records. It does not establish live performance.
Purchase orders · staff survey · NSSProvides a framework and ethical design questions. It is not a deployed product or measured model.
Yorùbá NLP · digital heritage archiveBEYOND MODEL ACCURACY
Evaluate the service people actually use.
My learning reports and catalogue connect technical evaluation with adoption, human agency and reliable operation. For a pilot, I would agree the baseline, review period and decision owner before collecting the following measures. This is an evaluation approach, not a claim that these outcomes have already been achieved.
| Question | Possible evidence | Decision it informs |
|---|---|---|
| Does the work improve? | Task completion and elapsed time against a comparable baseline, including review and rework. | Continue, revise the workflow or stop. |
| Can people exercise judgment? | Overrides, reasons for disagreement and whether explanations help reviewers. | Adjust autonomy and the human hand-off. |
| Who finds it difficult? | Completion, error patterns and feedback across relevant user groups and access needs. | Revise language, examples or accessibility. |
| Can the service be supported? | Data freshness, failed refreshes, recovery time and a named operational owner. | Repair reliability before extending access. |
| Does learning transfer? | A colleague applies a method to a new question and explains its limitations. | Improve guidance, practice or coaching. |
Read more: what belongs in the handover
The data-engineering catalogue emphasizes safe reruns, logs and monitoring. I would pair a repeatable analysis with its input definitions, quality checks, configuration, dependencies, run instructions and response to failure. A successful notebook run is one milestone; an owned and maintainable service needs more.
The social-context work adds a second check: a technically promising tool can still fail through poor access, unclear responsibility or a mismatch with users’ language and expectations. The cultural AI studies explore that question; the meeting-agent case shows why controlled checks and wider readiness remain separate.
Public synthesis of the v2/v3 MSc Learning & Impact Reports and the learning catalogue’s deployment, socio-technical systems and communication topics. Proposed measures, not retrospective performance statistics.
EXPLAINABLE AI & DATA / READ THE METHOD
Connect the learning to the evidence.
Use the categories to distinguish a capability, an algorithm, a Python package and a delivery step. A mention in a module or case study does not mean that capability is running on this page.
Follow the reasoning.
- QuestionChoose the decision to support
- MethodMatch the task to a suitable approach
- TestCompare with appropriate evidence
- ExplainState limits and human responsibility
Explore the building blocks.
Number of concepts in this guide, not performance or usage measures.
Explore the explanation matrix · 12 concepts
12 concepts shown
On small screens, scroll the table sideways. “Learning context” explains a related idea; it does not mean that method runs on this page. Deep learning is the standard term for “deep ML”.
| Technical term | Technical description / definition | In plain English | How this page uses it |
|---|---|---|---|
| Machine learning (ML)AI Capability | Methods that estimate patterns or parameters from examples to support inference on other inputs.Primary reference ↗ | Learn a pattern from examples, then check whether it helps with new ones. | Learning context ML learns from examples. Read the page-specific methods below to see whether any model runs here. |
| Deep learningAI Capability | Machine learning using neural networks with multiple representation-learning layers.Primary reference ↗ | A layered model learns useful representations; more layers do not guarantee a better answer. | Learning context Deep learning is an AI capability, not a label for every chart. No deep model is implied by this explanation panel. |
| Natural language processing (NLP)AI Capability | Computational methods for representing, analysing or generating human language.Primary reference ↗ | Help a computer work with words, while remembering that context matters. | Learning context Text search, translation, thematic classification and generation are different language tasks; their implementation is identified separately. |
| Sentiment analysisAI Capability | A language-analysis task that estimates an expressed attitude, often as polarity or sentiment categories.Primary reference ↗ | Estimate the tone of words—not the private feelings or personality of their author. | Learning context Sentiment estimates expressed tone. It must not be confused with topic classification or a diagnosis of a person. |
| Data science pipelineData Science Pipeline | An ordered, reproducible sequence for acquiring, validating, preparing, analysing and evaluating data.Primary reference ↗ | A chain from a question and source to an answer someone can check. | Method guide The learning-to-practice path connects framing, preparation, modelling, evaluation and communication. |
| Features and labelsAI/Data Concepts | Features are model inputs; labels are target values or categories used in supervised learning.Primary reference ↗ | What the model sees, and the answer it is learning to estimate. | Learning context Projects define relevant input features and any target labels before choosing an estimator. |
| Held-out evaluationEvaluation & XAI | Assessment on examples excluded from fitting and model selection, according to the declared split.Primary reference ↗ | Keep some answers out of practice so the test means something. | Learning context Unseen and appropriately separated cases help assess whether a learned pattern generalises. |
| Explanation versus causal proofEvaluation & XAI | A model explanation describes selected behaviour or contributions; it does not establish causal effects or full reliability.Primary reference ↗ | A reason for this output is useful, but it is not proof that the model is right. | Responsible use Explain what was actually done and what the available evidence supports. |
| NumPyPython Package | A Python library for multidimensional arrays and numerical operations.Primary reference ↗ | The numerical toolbox behind many Python calculations. | Learning context A numerical Python library; it is not an AI capability or an algorithm by itself. |
| pandasPython Package | A Python library for labelled tabular data manipulation and analysis.Primary reference ↗ | A programmable workbench for cleaning and comparing tables. | Learning context A tabular Python library; a useful preparation tool, not a prediction model. |
| scikit-learnPython Package | A Python library providing estimators, preprocessing tools and model-evaluation utilities.Primary reference ↗ | A toolkit of models and fairer ways to test them. | Learning context A Python library of algorithms and evaluation tools, distinct from each fitted model. |
| Human oversight and consentResponsible AI | People retain appropriate review authority, permission controls and correction routes for data use and outputs.Primary reference ↗ | Keep someone able to question, correct or stop the result. | Responsible use Decision ownership, permission and correction belong throughout the workflow. |
No matching concepts. Choose another category or clear the search.
