Turning complex businesses into structured solutions.

Team of five, with Giorgio Fiorentino, Gabriela Méndez, María Mora & Samreen Siddique · supervised by Carlos Farré Carrasco.

Do AI agents play the same way when the game changes? We had six large language models from Anthropic, OpenAI and Google play three classic games (an iterated Prisoner's Dilemma, a Commons Dilemma over a shared, regenerating resource, and a Cheap Talk signalling game) and measured each one on a five-parameter strategic profile: reciprocity, exploitation, honesty, trust and belief calibration.

• FingerprintsEach model has a distinct, reproducible strategic profile, and they separate clearly on reciprocity, exploitation and honesty.
• IncentivesWhen lying paid, four of six models cut truthful signalling sharply (a drop of 43 to 52 points); two barely changed.
• Structure over labelsHiding the state of the shared pool raised resource collapse from 37% to ~83% of games; telling a model its opponent was human or AI changed nothing.

Behind it is a lot of teamwork and a lot of analysis: designing the experiments, versioning every prompt, logging thousands of rounds into per-game databases, and testing the results with non-parametric tests, round-level models and robustness checks.

It was demanding, and one of the most satisfying pieces of work I've done. The lesson carries straight into business: incentives, information and the rules of the game shape behaviour far more than labels do.

Team project with Gabriela Méndez, María Angélica Mora Zamora, Georgii Runko & Samreen Siddique.

OpenTable curates, but it doesn't personalise: every diner sees the same "best of" lists, blind to the occasion. We designed a recommender that gets the table right when it matters most, like birthdays, anniversaries and business lunches, with one system that grows with each diner across three models:

• Cold user · 0 to 2 bookingsSegment-based collaborative filtering; a short onboarding survey solves the cold start.
• Warm user · 3 to 9 bookingsA hybrid of collaborative and content-based filtering, with the weights shifting automatically as the diner's history grows.
• Established · 10+ bookingsContext-aware matrix factorisation: occasion, party size and time re-rank the diner's own taste instead of replacing it with a generic list.

We tied it back to the business, with repeat booking rate as the primary KPI and booking conversion as secondary, and built the risks in from the start: a diversity quota against filter bubbles, re-ranking against popularity bias so new and niche venues still get seen, GDPR-minded use of personal data, and paid placement that can never override match quality. The demo runs the three models live on Barcelona restaurant data.

Academic project · not affiliated with OpenTable.