Two New Masters at Biorce

Research that matters

These theses weren't academic exercises on the side, they tackled real challenges in clinical trial development, the kind Biorce works on every day.

Marcelo's thesis, "Automatic Prompt Optimisation for Patient-to-Clinical-Trial Matching", is a great example. It addresses one of the biggest bottlenecks in medical research: patient recruitment. Consider the scale of the problem:

• Only 7.1% of adult cancer patients enrol in a treatment trial.

• Manual prescreening takes 3.4-8.8 staff-hours per enrolled patient.

• Insufficient recruitment is the most common reason trials are discontinued.

His work introduces Celestine, a database-agnostic AI framework that matches patients to trials, and Capo, a method that automatically optimises the system's prompts instead of hand-crafting them. The results speak for themselves: Capo raised the system's reliability score from 0.50 to 0.62, and its gains held up even on a database dialect never seen during development, a key step towards AI systems that hospitals can actually trust and reuse.

Philipp's thesis, "With Great Trust Comes Great Adoption: Reliable Confidence Metrics for Short-Form LLM Generations in a Post-Hoc Setting", tackles the question every team deploying LLMs eventually hits: when the model gives you an answer, can you trust it? His experiments show that, out of the box, the answer is no:

• Confidence scores reported directly by the model are systematically too high and, when interpreted as probabilities, provide little information beyond the model’s overall accuracy.

• No single confidence signal works in every setting. The right one depends on how the model is run, and choosing the wrong signal can cause it to fail silently.

• Miscalibration has domain-specific shape: it hides in aggregate statistics and concentrates exactly on the classes where errors matter most.

His work delivers a complete, provider-agnostic methodology that fixes this by turning a single API call into confidence you can act on: it selects the right confidence signal for the setup at hand, and repairs its probabilities with a cheap, lightweight calibration step learned from a handful of labelled examples. The payoff is a new way to run human-in-the-loop review: stakeholders set the acceptable error rate up front, with a statistical guarantee attached, and the system derives how much can be safely automated and routes the rest to human experts. This methodology is a key step towards LLM automation that stakeholders can trust and audit.