PhD researcher in Medical Sciences (AI in Oncology) at the University of Cambridge, working at the Cancer Research UK Cambridge Institute.
I design multimodal machine learning systems that integrate histopathology, genomics, and clinical data for early cancer progression prediction, with emphasis on reproducibility, safety, and deployment in research infrastructure.
rdaz2@cam.ac.uk · LinkedIn · Schedule a call
Developing and benchmarking multimodal models combining histopathology and copy number variation data to predict progression risk in Barrett’s oesophagus.
Compact benchmark for testing prompt injection, answer poisoning, protected-context leakage, and unsafe tool use in retrieval-augmented generation systems, with a minimal end-to-end RAG pipeline, lightweight defenses, and reproducible evaluation reports.
Compact benchmark for testing whether simple monitors can catch suspicious coding-agent behaviour on realistic toy software tasks, with structured action, diff, test, and timing logs plus clean evaluation and failure-case reporting.
Lightweight CLI for turning predictions, metrics, configuration, and optional logs into standardized evaluation reports with markdown, HTML, and machine-readable summaries across classification, ranking, and agent tasks.
Privacy-aware early public-health signal monitoring prototype using synthetic data, interpretable anomaly detection, and cautious research-first framing.
Building a SLURM-backed local LLM inference tool for CRUK CI that enables reproducible, logged, GPU-based model execution on institutional HPC infrastructure.
Designing and running repeated public-goods games where agents are LLMs executed as SLURM jobs via Ollama on HPC infrastructure. The framework enables controlled experiments on multi-agent coordination, contribution dynamics, communication effects, and model-level behavioural differences, with full job-level logging and experiment aggregation.
Minimal macOS SwiftUI application for real-time voice processing using AVAudioEngine. Implements live autotune intensity control, reverb control, device routing, and record mode with timestamped audio export in a lightweight local pipeline runnable via Swift Package Manager.
Chrome extension for blocking distracting websites with a local dashboard, built-in social and adult/NSFW groups, custom domain rules, and recurring weekly schedules.
Electron desktop app for spawning and viewing multiple live terminal windows in one place, intended as an early foundation for a more visual multi-session workflow.
Weekend flight tracker for Geneva and London airports with a lightweight TypeScript workflow for route-level monitoring.
Static GitHub Pages website for Life in the UK practice questions with a fast topic-first workflow and pre-computed LLM context generation for each question.
Terminal-first Spanish cloze practice over SSH with spaced repetition, local progress tracking, and sentence cards built from the Tatoeba English-Spanish corpus.
Cell Segmentation-Classification-Tracking