AI defect recognition approved for submarine-construction shipyards
Took a radiographic defect-recognition system from a blank page to DoD approval for shipyard integration, anchoring roughly $7M in consecutive contracts.
- three consecutive DoD contracts
- ~$7M
- approval for shipyard integration
- DoD
- Organization
- ORBIS, Inc. · DoD program
- My role
- Director of Artificial Intelligence; program lead
- Period · Focus
- 2020 – 2026 · Platforms, Strategy, Governance
In brief
Radiographic inspection of welds and castings decides whether a flaw ends up inside a submarine hull. I procured and led three consecutive DoD contracts, about $7M in total, for an AI system that recognizes those defects, built and optimized its training cycle from scratch inside the accredited boundary, and secured copyright on the algorithm. The system earned Department of Defense approval for shipyard integration on submarine builds and drew inbound requests for commercial deployment.
Context and challenge
Inspectors review radiographic imagery of welds and castings where a missed defect has physical consequences. A model had to earn trust from regulators and from the inspectors who could simply route around it, and the imagery could not leave a controlled environment.
My mandate and scope
- Procured and led three consecutive DoD contracts worth approximately $7M.
- Built and optimized the defect-recognition training and experimentation cycle from scratch, on controlled radiographic imagery inside the accredited boundary.
- Led integration into the shipyard inspection workflow and the evidence package behind DoD approval.
Key decisions and why
- 1
Build the evaluation harness before the model matured
In defense work the harness is reviewed as closely as the model. Performance had to be demonstrated repeatably across defect types, imaging conditions, and edge cases. That discipline cost real budget and became the program's main advantage: documentation answered stakeholder challenges, and every model change reported what it cost.
- 2
Design the inspector hand-off as carefully as the network
The system changed what inspectors spent their attention on; it did not replace them. Which findings surface, how confidence is presented, and where a human makes the call took as much engineering judgment as the architecture, because an inspector who distrusts a tool works around it and the accuracy metrics stop mattering.
- 3
Train inside the boundary instead of cloud-training and importing
The data could not leave, so the full training and experimentation cycle had to run on-premises. Every capability had to be justified, built, and maintained inside the boundary, and several of the better architectural decisions traced back to that constraint.
What changed
- Inspection attention shifted toward the findings the system surfaced, with a human retaining the decision.
- Program budget was weighted toward evaluation, integration, and the people operating the result, since that is where the program's risk lived.
Outcome and evidence
- Department of Defense approval for shipyard integration on submarine builds.
- Three consecutive DoD contracts, approximately $7M combined.
- Inbound requests from commercial organizations seeking deployment partnerships.
What to expect from me
I treat evaluation and the human workflow as the product, with the model as one component. In a regulated or high-consequence setting I will put money into evidence and adoption before chasing another point of accuracy.
Technical perspective
- Computer vision and image segmentation on radiographic imagery of weld and casting defects; the core recognition algorithm is copyrighted company IP.
- Training and experimentation ran entirely on-premises on controlled data; deployment followed the consuming inspection workflow.
- Evaluation covered defect types, imaging conditions, and edge cases, with documented, repeatable results for each model change.
- Operational risks considered before development included input drift from new imaging equipment, reviewer workload, and the evidence a DoD reviewer would require.
Artifacts
Confidentiality: Imagery, model architecture, performance figures, and program specifics are controlled and not published. Only the publicly shareable outcomes appear here.