Senior Evaluation Specialist | Research Fellow
Lucas Sempé
I am a Brazilian-Argentine researcher working on how inequality shapes health, learning and care, mainly in Latin America and sub-Saharan Africa. Much of my empirical work asks how far public services reach the people they are meant to: whether health financing arrives at primary care facilities, how wealth inequality within schools relates to learning achievement in PISA, and what community-based care for older people changes in Brazilian health services. The same question runs through my work on mental health — from the relationship between neglected tropical diseases and distress in Liberia to social media use and self-harm among secondary school students in England — and through work on AI, where I study governance capacity across 193 countries and persona drift in mental-health chatbots by reading model activations. Methodologically I work on causal inference when evaluations have few cases, and on evidence synthesis across health and social protection. Before returning to research I directed Peru's performance-based education budget, at £2 billion and 55% of national education spending; the gap between the evidence we had and what the budget cycle rewarded is what led me back.
Books
Open, practical books
Machine learning
Machine Learning Concepts
A four-level handbook from intuition and mechanics through mathematics, code, research, and systems.
Impact evaluation
Small-n Approaches to Impact Evaluation
A guide to selecting, combining and assessing evaluation approaches when conventional large-sample designs are not appropriate.
Recent Posts
View all →
The Instruction Tuning Firewall
Mental health chatbots can drift toward dangerous validation while sounding perfectly appropriate. I built a monitoring system that detects persona drift in model activations—catching problems that even a fine-tuned DeBERTa misses, with a 2.6× advantage on crisis recognition. Validated by two clinical psychologists (ICC=0.716) and tested on naturalistic emotional support conversations.
When Algorithms Meet Warzones
A drone image classifier that can't distinguish combatants from farmers. A beneficiary targeting model trained on data from before the displacement. A chatbot collecting trauma narratives in a language it barely understands. These aren't hypotheticals—they're the edge cases where AI meets impact evaluation in fragile contexts.
The Capacity Gap
I scored 2,216 AI policy documents across 193 countries on implementation capacity. The headline isn't that rich countries do better—it's that the gap nearly vanishes once you account for documentation quality. The real story is what's happening within income groups.
Talking to Your Evidence Base
What if you could ask your research library a question out loud and get a spoken answer grounded in actual studies? A retrieval-augmented system with voice interface makes research synthesis conversational.