← Applied AI engineering
Working prototype
EvidenceGate
Controls around model output
Architecture and implementation
2026
- Python
- OpenAI Responses API
- Citation validation
- Human approval
The problem
A plausible model finding must not become an approval simply because it is well written. The system needs to validate citations and keep provider failures visible.
What I did
- Separate deterministic checks from optional model analysis.
- Reject unsupported citations and malformed findings rather than treating them as acceptable output.
- Keep approval, merge and deployment authority outside the model.
The result
19 automated tests and four synthetic scenarios exercise the prototype controls in the CV evidence snapshot.
Scope and boundaries
A working prototype with synthetic scenarios. It is not presented as sustained production operation or a compliance certification.
Portfolio summary updated September 2026.