Miriam Sanguedolce

I've built and locally deployed an LLM-assisted agent for my MSc thesis at the University of Limerick, combining Qwen2.5-3B-Instruct with deterministic analytical tools to explain validated haemodialysis modelling results.

I designed controlled prompts, structured JSON inputs and outputs, validation, safeguards, fallback behaviour, and JSONL audit logging. My reproducible workflow covered 2,276 synthetic longitudinal records, causal and predictive modelling, five-fold cross-fitting, and uncertainty analysis.

At RWS Group, I evaluate search and AI-generated outputs for relevance, accuracy, usefulness, and consistency. I also lead a distributed moderation and support team at StudyStream, improving quality standards, workflows, internal tools, and product feedback processes.