Capstone: Visual Weld-Defect Inspection

The capstone runs the length of the program. In Agile sprints, your team builds an image-analysis model that flags weld defects from photographs — applying each module's skills to a real manufacturing-inspection problem. Tuesdays transition from labs into project work; Wednesdays are facilitated project days.

Goal: a working prototype that classifies weld images (e.g., acceptable vs. defective), backed by evidence it is trustworthy — good metrics on unseen data and interpretability showing it focuses on the weld, not the background. See the 12-week roadmap for how modules sequence.

Sprint plan

  1. Weeks 1–2

    Sprint 1 — Frame the problem

    Modules 1–2

    Stand up an Agile team and board. Write a problem charter: what defect are we detecting, why it matters, and how we measure success. Assess it honestly as a data science problem and note ethical/safety concerns.

    Deliverable: Problem charter + success metric
  2. Weeks 3–5

    Sprint 2 — Data foundation

    Modules 3–4

    Set up the dev environment and git workflow. Ingest weld/inspection images, organize them into labeled folders, and clean/validate the dataset for training.

    Deliverable: Versioned dataset + data card
  3. Weeks 6–7

    Sprint 3 — Explore & engineer

    Modules 5–6

    Explore the image set: class balance, lighting, resolution. Build augmentation and preprocessing pipelines and prepare train/validation/test splits without leakage.

    Deliverable: EDA notebook + preprocessing pipeline
  4. Weeks 8–9

    Sprint 4 — Train

    Module 7

    Establish a simple baseline, then train a CNN (including transfer learning). Tune hyperparameters and use GPUs/batching to train efficiently.

    Deliverable: Trained model + training report
  5. Week 10

    Sprint 5 — Evaluate & explain

    Module 8

    Measure precision, recall, and F1 (accuracy alone misleads on rare defects). Do error analysis and use Grad-CAM to confirm the model looks at the right regions.

    Deliverable: Evaluation + interpretability report
  6. Weeks 11–12

    Sprint 6 — Deploy & present

    Module 9

    Package the model behind a simple inference path, write a monitoring and retraining plan, and present the end-to-end solution to stakeholders.

    Deliverable: Deployment plan + final demo

Deliverables

  • Problem charter with a measurable success metric
  • Labeled image dataset with a data card
  • Reproducible training pipeline in git
  • Trained model with precision/recall/F1 on a held-out set
  • Interpretability report (Grad-CAM) and error analysis
  • Deployment and monitoring plan
  • Final stakeholder presentation

Suggested team roles

Product owner Owns the problem statement and priorities; speaks for the stakeholder.
Scrum lead Runs stand-ups and the board; removes blockers.
Data lead Owns dataset quality, labeling, and the data card.
Modeling lead Owns training, tuning, and evaluation.