Automated Defect Detection Model
FY26 · 57 evidence items · 57% claim readiness
Project summary
FlowForge is building a computer vision model to detect inconsistent surface defects under variable lighting and material conditions, removing the need for manual visual inspection on the production line.
Technical uncertainty
The team could not determine whether a convolutional model could reliably distinguish genuine surface defects from acceptable surface variation across the range of material types and lighting configurations present on the production line.
Working hypothesis
“If a computer vision model were trained on labelled defect images across multiple material classes, it may maintain acceptable precision and recall under the variable lighting conditions present in the production environment.”
Claim readiness
Key figures
- Est. eligible spend
- $146K
- Evidence items
- 57
- Confidence
- Medium
- Documentation gaps
- 2
Evidence gaps
- Missing model comparison report
- Unclear cost allocation
Experimentation timeline
Jan 2026
Defect detection model project initiated
JiraFeb 2026
Initial model evaluation — insufficient precision
DriveMar 2026
Dataset augmentation and retraining commenced
GitHubApr 2026
V2 model tested — false positive rate too high
JiraMay 2026
Custom CNN with lighting normalisation in progress
GitHub
Experimentation iterations
Off-the-shelf defect classifier
Outcome
Insufficient precision on variable surface textures
Evidence
Model Eval V1 — Google Drive
Fine-tuned ResNet with augmented dataset
Outcome
Improved recall but high false positive rate under low light
Evidence
Model Eval V2, Jira DEF-74
Custom CNN with lighting normalisation layer
Outcome
Ongoing — performance under evaluation
Evidence
Model Eval V3, GitHub PR #301
Supporting evidence (2 items)
View all evidenceTeam involvement
- Maya ChenML Engineer65% R&D time42 signals
- Anika PatelManufacturing Lead45% R&D time28 signals
- Liam BrooksProduct Engineer28% R&D time19 signals
Eligible cost breakdown
$146K
Total estimated eligible expenditure
Confidence breakdown
Missing information
View all gaps- Upload model comparison report showing performance across versions
- Clarify cost allocation between ML compute and routine operations
- Link dataset acquisition invoices to project activity
- Confirm team involvement percentages for ML engineers
Figures update automatically as new evidence is connected. Your Canopy specialist will confirm all positions during claim preparation.