About the role
You help a private agent learn what matters to its user while preserving their control and their right to change their mind. The distinction that governs the work: memorisation is not personalisation, and a system that cannot forget is not one a person actually controls.
The work
Research adaptation, memory selection, preference learning and resistance to forgetting or poisoning. Compare retrieval, explicit settings and parameter updates rather than assuming training is always necessary. Test what can be removed from memory and what model-level removal can actually guarantee.
What good looks like
In your first 90 days, deliver a personalization experiment with held-out evaluation, rollback and documented deletion limitations.
Evidence we look for
Bring machine-learning research expertise and careful reasoning about privacy, causality and evaluation. Show work that distinguishes memorization from useful generalization.
What we need to see
- Machine-learning research expertise with careful reasoning about privacy, causality and evaluation
- Work that distinguishes memorisation from useful generalisation, which you can walk through
- You design for forgetting and revision as first-class operations
- You can evaluate personalisation without a large pooled dataset
Nice to have
- Continual learning or catastrophic forgetting research
- Causal inference
- Recommender systems with a privacy constraint
The exercise
Design an experiment where a user's preferences change and the system must update without exposing another household member's data.
Where and how we work
In the office together five days a week, in any of these cities. Remote-friendly around your family, arranged one person at a time.