I test AI systems and document where they fail. Across four hands-on audits, I evaluated hallucination, prompt injection, bias, and EU AI Act compliance using predefined scoring rules and repeat runs, reporting null results honestly.
For my dissertation, “Understanding Human Bias In Multimodal AI Systems (SigLIP 2 and 3 VLMs),” I designed a counterfactual gender-bias test and measured embedding and behavioural bias across vision-language models. I built the analysis pipeline in Python and corrected four errors in my own code before rerunning the analysis.
In my ChatGPT customer-support chatbot audit, I created a 50-case test kit and tested identity swaps. I rewrote the system instructions, fixing all 11 failures with no regressions; answers became identical across 20 of 20 runs, compared with 5 of 20 before the fix.
I also assessed an AI CV screener against the EU AI Act, tested prompt-injection protections, and built an Excel monitoring sheet with impact-ratio alerts. In my ongoing CV-screening project, I built an offline browser dashboard that flags hidden content and sends it for human review rather than automatic rejection.
