Scenario Sprint: pick the technique, defend the call

A team sprint — 3–5 people, one 60-minute block. Each team takes one client, frames the problem, chooses and tests a technique, and defends the recommendation in a three-minute pitch. There is rarely one right answer; there is always a justified one.

60:00 Suggested rhythm: 45 min team work → 15 min pitches (3 min each + 1 min questions).

Pick your client

Each team selects one scenario. Multiple teams on the same scenario makes a better comparison panel at pitch time.

🏥 Melbourne Health Network — patient readmission prediction

"We need to identify patients at high risk of readmission within 30 days. Our current process is inconsistent — some doctors spot at-risk patients, others miss obvious cases. We want AI to help, but any recommendation must be explainable to medical staff and auditable for compliance." — Dr. Sarah Chen, Chief Medical Officer. Stakeholders: medical staff, hospital administrators, regulators, patients.

Available data (2 years, 50,000 records)

  • Demographics: age, gender, location, insurance
  • Clinical: diagnosis, comorbidities, severity scores
  • Treatment: length of stay, procedures, medications
  • History: previous admissions, compliance
  • Outcome: readmitted within 30 days (yes/no) — 12% base rate

Team tasks (45 min)

  • Problem analysis (10): classification or regression? Which business requirement dominates?
  • Technique selection (15): compare decision trees against alternatives
  • Build & test (15): implement the recommended approach
  • Business case (5): recommendation plus justification

Workflow starters

[File] → [Data Sampler] → [Tree] → [Tree Viewer] + [Test & Score]

[File] → [Data Sampler] → [Tree, SVM, kNN] → [Test & Score]

Widgets worth exploring: Tree Viewer (interpretable rules), Test & Score, Confusion Matrix, Feature Statistics.

Pitch deliverables

  • Technique recommendation — which approach, and why
  • Explainability strategy — how doctors read the recommendations
  • Performance metrics — accuracy, precision, recall
  • Implementation plan — integration into clinical workflows
  • Risk assessment — what could go wrong, and mitigation

Debrief: the lesson is the tension between accuracy and interpretability under regulation — not the winning model. Cross-link: technique-picker, scenario 2.

💳 Aussie Fintech Startup — credit risk assessment

"We're disrupting traditional lending by serving customers the banks reject. We need AI more sophisticated than a credit score, but we can't build a system that discriminates against protected groups — and we must explain our decisions to regulators." — Alex Thompson, CTO. Challenges: serve the underbanked fairly; compete on speed; comply with anti-discrimination law; keep the AI advantage.

Available data (18 months, 25,000 applications)

  • Traditional: credit score, income, employment history
  • Alternative: utility payments, rent history, education
  • Behavioural: application completion patterns, response times
  • Social: references, guarantor information
  • Outcome: default within 12 months (yes/no) — 8% base rate

Team tasks (45 min)

  • Problem analysis (10): classification — but what are the fairness implications?
  • Technique comparison (15): accuracy and bias across approaches
  • Build & analyse (15): feature importance and demographic correlations
  • Ethical assessment (5): fairness and regulatory compliance

Workflow starters

[File] → [Data Sampler] → [Tree, SVM, kNN] → [Test & Score] + [Confusion Matrix]

[File] → [Rank] → [Data Table]

[File] → [Select Columns] → [Scatter Plot]

Use the scatter plot to check score behaviour across demographic groups — the bias check matters more than the accuracy number.

Pitch deliverables

  • Technique recommendation for accuracy and fairness
  • Bias analysis — how protected groups are protected
  • Performance comparison across techniques
  • Regulatory compliance story
  • Competitive advantage argument

Debrief: an 8% default base rate means "always approve" scores 92% — the accuracy headline is a trap. Connect to threshold-dial and k-anonymity.

🛒 National Supermarket Chain — dynamic pricing optimisation

"We want AI-driven dynamic pricing like the online giants, but for groceries. Prices must respond to demand, competition, inventory, and customer segments — and our pricing is highly visible to customers and competitors. It has to be effective and defensible." — Maria Rodriguez, Head of Pricing Strategy. Requirements: real-time recommendations for 50,000+ products; customer trust; maximised margin.

Available data (3 years, millions of transactions)

  • Sales: product, store, time, price point
  • Inventory: stock levels, shelf life, supplier costs
  • Competitor: rival pricing, market share
  • Customer: loyalty data, purchase patterns
  • External: weather, events, economic indicators
  • Target: optimal price (continuous) or price tier (low/med/high)

Team tasks (45 min)

  • Problem framing (10): regression (price) vs classification (price tier) — which serves the business?
  • Technique testing (15): compare with real-time requirements in mind
  • Build & evaluate (15): alternative pricing models
  • Deployment strategy (5): rollout across 1,000+ stores

Workflow starters

[File] → [Data Sampler] → [Linear Regression] → [Predictions] + [Test & Score]

[File] → [Data Sampler] → [Tree] → [Tree Viewer] + [Test & Score]

[File] → [Data Sampler] → [Random Forest] → [Test & Score]

Pitch deliverables

  • Problem-type decision: regression vs classification
  • Technique recommendation for real-time pricing
  • Performance analysis with speed considered
  • Scaling strategy for a national chain
  • Customer-trust impact

Debrief: the framing decision (continuous price vs price tier) changes everything downstream — most teams never notice they chose it. Connect to technique-picker, scenario 1.

⛏ Australian Mining Corp — equipment failure prediction

"Equipment failure in our mines isn't just expensive — it's potentially deadly. We need AI that predicts failures before they happen, but we can't afford false alarms that shut operations down unnecessarily. Predictions must be accurate, timely, and explainable to the maintenance teams acting on them." — David Kim, Operations Director. Requirements: catch real failures; control false positives; explainable; reliable in harsh conditions.

Available data (5 years, 200+ machines)

  • Sensors: temperature, vibration, pressure, electrical
  • Maintenance: service history, part replacements, downtime
  • Environment: weather, operational intensity
  • Fleet: age, manufacturer, model, usage patterns
  • Outcome: failure within next 7 days (yes/no) — 3% base rate

Team tasks (45 min)

  • Problem analysis (10): classification with extreme cost asymmetry — how do you weigh missed failures vs false alarms?
  • Technique evaluation (15): reliability and explainability first
  • Test (15): build models, read the confusion matrix carefully
  • Safety strategy (5): AI augmenting human judgement, not replacing it

Workflow starters

[File] → [Data Sampler] → [Tree] → [Tree Viewer] + [Confusion Matrix]

[File] → [Data Sampler] → [Random Forest] → [Test & Score] + [Confusion Matrix]

The confusion matrix is the whole assessment: false negatives are dangerous, false positives are expensive. A 97% accuracy headline can hide every failure.

Pitch deliverables

  • Technique selection: accuracy vs explainability balance
  • Safety analysis: minimise false negatives, control false positives
  • Confusion matrix analysis with error costs
  • How maintenance teams use the predictions
  • Human–AI interaction: augmentation, not replacement

Debrief: a 3% failure rate makes this threshold-dial's argument with lives attached — 97% accuracy by predicting "never fail" hides every real failure. Also ask: what's missing from five years of "complete" records? (Compare technique-picker, scenario 5.)

Pitches

Each team: 3 minutes presentation + 1 minute questions.

Pitch structure

  • 30s — the business problem
  • 60s — the technique decision and why
  • 60s — evidence: workflows and results
  • 30s — business impact

What good looks like

  • Clear business logic — why this technique fits this problem
  • Trade-off analysis — accuracy vs interpretability vs implementation
  • Evidence — workflows and results that support the call
  • Professional communication — technical concepts for business stakeholders

Debrief points

  • Different business contexts demand different techniques — there is no default
  • Rarely one correct answer; the defence is the deliverable
  • Connect forward: k-means-stepper (clustering), perceptron (mechanics), threshold-dial (error costs)
  • The durable skill: requirement analysis and technology selection, explained to non-technical stakeholders