
Sep 2025 - May 2026130+ASSESSMENT CRITERIA
5ASSESSMENT SECTIONS
3EVIDENCE FILE TYPES
16+RAG DOCS
Built the backend of the NCSC Certified Degree Application tool, helping assessors grade university cyber-security degrees against 130+ criteria
- Integrated the OpenAI API into live grading code, each call using a section-scoped prompt, a fixed evidence boundary and a required response schema so results came back in a dependable shape.
- Built RAG retrieval over PDF evidence plus targeted extraction from Word and Excel, so the model only ever saw the passages tied to the criterion being graded, cutting tokens, latency and making every verdict traceable.
- Split grading between deterministic rules and the model: objective checks ran as code and the LLM handled only the criteria needing judgement, removing a whole class of avoidable errors.
- Kept a human in control, with every verdict returned as an editable draft and nothing final until an assessor saved it and a panel confirmed it at consensus stage.
- Tested the workflow across 8 cases covering complete, incomplete, miscategorised and near-empty evidence, including the ones meant to fail and stop the workflow.
- Received a letter of appreciation from the NCSC Deputy Director for National Resilience Capabilities.
PythonFastAPIPostgreSQLRedisSQLAlchemyOpenAIRAGVector SearchReportLab