Researcher working on information markets, mechanism design, and multi-agent systems.
I work at Gensyn on information markets: mechanisms that pay people and models to reveal what they know, and the multi-agent systems that take shape once many strategic participants trade on the result. Most of my work sits where mechanism design, information elicitation, and multi-agent learning intersect.
Before Gensyn I was an Applied Scientist at Amazon, working on fraud and bad-actor detection. I received my PhD in computer science from the University of Colorado Boulder in 2022, advised by Rafael Frongillo and Joshua Grochow, where I studied the dynamics of learning in games. Before that, I did an MS in applied mathematics at UMass Amherst.
Order-book market makers in prediction markets do not earn from spreads but from carrying the under-priced side of behavioural tail demand to settlement, which screens informed traders out of the order-book and towards any coexisting automated market maker.
If the first-pass report in a dispute-based oracle is credibly neutral and accurate, the vote needed to overturn it can be raised and manipulation becomes provably harder; without neutrality, no threshold does better than a coin flip.
A generalization of prediction markets that pays traders for evidence as well as forecasts, and can resolve on the accumulated evidence when no external ground truth exists.
A two-sided market design for goods that lose their value if unused, such as compute. We use a load-based posted price to handle price discovery and a greedy second-price matching rule to handle allocations, so providers are kept truthful.
A fully decentralised, asynchronous algorithm for post-training language models with reinforcement learning, in which heterogeneous nodes share rollouts rather than weights so that useful behaviour spreads across the network.
Ensembles of small domain-expert language models trained in parallel do better when the experts differ in size and training time to suit their domain, beating homogeneous ensembles on 20 of 21 domains at the same compute budget.
Replicator dynamics in matrix games can simulate any Turing machine, so basic questions about where learning in a game ends up are undecidable.
Tools for understanding multi-agent economic systems away from equilibrium: learning in games is shown to be universal from the viewpoints of both dynamical systems and computability, alongside graphical models of exchange and algorithms by which decentralised agents reach approximate equilibria.
Learning dynamics in matrix games can approximate any dynamical system, chaotic ones included.
Electricity markets modelled as smooth games whose consumers are gradient-based learners, with examples of how instabilities arise even under reasonable local objectives and conditions under which local Nash equilibria are stable fixed points of the dynamics.
Exchange economies on a network, where agents trade only with their neighbours, extended to allow resale, and what that does to equilibrium prices.
Network models of oscillatory neural dynamics for associative memory.