Technology Feasibility Summary

1. Vision transformers for defect detection

Overview: Replace CNN classifier with a ViT fine-tuned on defect imagery.
- ✅ Pros: higher accuracy on subtle defects; better transfer learning.
- ⚠️ Cons: needs more labelled data and GPU for training.
- Organisational fit: M — team has ML skills, limited GPU budget.
- Recommended action: Pilot on a 2k-image subset; decide on accuracy delta vs CNN.

2. Edge inference on Jetson

Overview: Run inference on-device to cut latency and cloud cost.
- ✅ Pros: <50ms latency; offline capable; lower opex.
- ⚠️ Cons: model quantisation effort; hardware procurement lead time.
- Organisational fit: H — aligns with existing on-prem deployments.
- Recommended action: Proceed — quantise current model, benchmark on 1 unit.

3. RAG over internal docs

Overview: Retrieval-augmented assistant over internal knowledge base.
- ✅ Pros: fast to ship; immediate productivity gain; low risk.
- ⚠️ Cons: data governance; needs eval harness to avoid hallucination.
- Organisational fit: H — clear ROI, low infra.
- Recommended action: Proceed — start with a 2-week scoped pilot on one team.

Decision-ready. Confirm data availability and budget before committing.