Search published Practitioner Experience Notes (PENs). Members can view any PEN; only the author and admins can edit. Authors can choose to display their name or remain anonymous to Members, though public search results will always be made anonymous.
Who or what is Eglantyne? What sort of questions can she answer?
Eglantyne is an AI Research Assistant built to give trusted answers to a broad range of humanitarian questions. The breadth of question is deliberately exceptionally large within the confines of humanitarian matters; current affairs and programmes, history, people, organisations, trends, techniques, advice … there are no restrictions within the wider remit.
The principle design considerations were around trust, and to build a system which could be widely delivered free at the point of usage.
Eglantyne uses powerful AI inside a controlled evidence-and-assurance process and is essentially a retrieval-augmented AI system built around OpenAI language models.
When asked a question, Eglantyne does not simply send that question to an OpenAI model and accept whatever comes back. It first interprets what is being asked and determines the type and breadth of research required. It then searches the knowledge sources available to it — including curated humanitarian material and the Alumni Association’s PENs — and retrieves the material most relevant to that particular question.
That selected evidence is then supplied to the OpenAI model with detailed instructions about how it should analyse it, what weight to give different kinds of evidence, how to handle conflicting or incomplete information, and how the answer should be structured. The model’s job is therefore principally to reason over retrieved evidence, synthesise it and explain it, rather than to rely on what it happens to remember from its original training.
Eglantyne then presents the result with its sources and limitations to expose the evidential basis for the answer.
‘Trusted’ does not mean that the AI has somehow been made incapable of making mistakes. Large language models are probabilistic systems and can be wrong. Eglantyne instead tries to create trust at the system level: constrain the model with relevant evidence, favour authoritative and directly applicable sources, make provenance visible, expose uncertainty, test retrieval and answer quality systematically, and continually test changes against a growing suite of known questions.
Eglantyne is under a continuous development programme in the fast-moving AI world eg it tries to optimise usage an array of the latest OpenAI models. It is currently working towards operating in two modes; Conversational and Essay.
Eglantyne was conceived and developed by a group of Save the Children Alumni for the benefit of the wider community. Pete Smith was Programme Director who both envisioned the product and built it. Mike Aaronson provided considerable governance direction and testing assessments. Leonie Lonton named Eglantyne in honour of the founder of the Save the Children movement, Eglantyne Jebb.