I research the evaluation and safety of clinical AI: whether the checks we run on AI systems actually catch the failures they're meant to catch.
Based in Mauritius. Applying to PhD programmes for 2027.
- Stress-testing sandbagging detection under adaptive attack. With two small training-data edits, a text-audit detector's AUC fell from 0.978 to 0.560, and activation probes flagged benign fine-tuning as sandbagging. Accepted, ARRSI 2026 (AACL-IJCNLP workshop). → ai-sandbagging
- When internal validation misses construct mismatch in multimodal mental-health ML: a documented failure case and three diagnostics to catch it. Under review, JMIR Mental Health. → multimodal-BD-Detection
- Multi-agent medical scribe: led the evaluation of LLM-generated SOAP notes with LLM-as-a-Judge. Write-up
- OpenMeet: private meeting transcription and summarisation on macOS, with Whisper.cpp and local LLMs. Nothing leaves your machine.