WhizStart

Train one generative sequence model over raw player event streams (bets, sessions, deposits, game plays) so it yields reusable representations for many downstream tasks and can generate plausible future behaviour. The research is in the architecture: behaviour lives at three time-scales at once, sequences are very long, and events carry continuous typed attributes rather than words. Includes a small scaling study.

Skills & interests: Deep learning, sequence modelling, generative models; some exposure to long-context or state-space architectures.

Find, without labels, a small set of interpretable player archetypes — sports-led vs casino-led, recreational vs high-intensity, discovery-seeking vs habitual — and show they are real, stable over time and useful downstream rather than clustering artefacts. Includes how players move between archetypes over their lifecycle.

Skills & interests: Unsupervised and representation learning; interest in interpretability.

Recommenders that optimise clicks concentrate players on a few familiar games and starve new titles. Reformulate recommendation as an explicit trade-off between immediate engagement, long-term player value and discovery, produce a controllable Pareto frontier, and test whether optimising for long-term value actually changes what gets recommended.

Skills & interests: Recommender systems and optimisation.

Treat recommendation as a sequential decision with principled exploration, and — the real contribution — work out when a new policy's value can be estimated reliably from logged data before it is ever shown to a real player. Produces an evaluation toolkit the whole team reuses.

Skills & interests: Reinforcement learning or online learning; statistics.

Today's bonus changes tomorrow's behaviour, so the best sequence of actions is not the best single action. Learn a sequential policy purely from logged decision data using conservative offline RL methods, and give an honest account of where offline RL can and cannot be trusted on real data.

Skills & interests: Reinforcement learning, especially offline or batch RL.

Given an estimate of how much each action would help each player, a fixed monthly budget, contact-frequency caps and eligibility rules, allocate finite interventions across the whole player base. Compare greedy, linear-programming and uncertainty-aware allocation on realised value.

Skills & interests: Operations research or optimisation, with some ML.

Build a world model of how players respond to actions, faithful enough to rank recommendation and intervention policies offline, and design the validation protocol that shows its verdicts predict real outcomes. The highest-leverage project in the portfolio: every other team iterates faster if the simulator can be trusted.

Skills & interests: Generative modelling, model-based RL, sequence modelling.

Build a small team of agents — a strategist that proposes campaigns, an evaluator that tests them against the simulator, a commercial critic that can veto — and test rigorously whether structured critique produces better strategies than a single agent or an analyst alone, including red-teaming the critic with deliberately weak proposals. A human approves everything.

Skills & interests: LLMs and agent systems; interest in evaluation.