More selected work
Side projects spanning multi-agent learning, generative audio, and computational humanities.
Multi-Agent Prisoner's Dilemma Dynamics
Trained populations of agents in parallel via gradient descent on reward, with memory of past interactions — producing emergent gradient dynamics over strategies.
Result: cooperative coalitions periodically emerged, but the system most often collapsed to universal defection. This convergence to a globally non-optimal equilibrium is consistent with running a local optimizer (gradient descent) over decentralized agents — each agent climbs its own reward landscape with no mechanism to escape the socially suboptimal basin.
Music Transformer for Genre Interpolation
Trained a transformer on instrumental scores to generate new pieces and interpolate between musical genres in latent space.
Result: the model produced coherent instrumental scores and smooth genre crossfades — interpolating along latent directions yielded pieces that audibly drifted from one genre's rhythmic and harmonic style to another rather than collapsing to a single mode.
Code →Semantic Search & Manifolds of Ancient Texts
Built a contextual topic-similarity search over ancient texts using sentence transformers, then analyzed the embedding manifold across corpora.
Result: embeddings clustered by theme rather than by source text, revealing low-dimensional structure that connected passages across different works on shared topics — and the manifold's geometry exposed which corpora occupied overlapping vs. distinct regions of conceptual space. The same embeddings were also discriminative enough to identify authorship on their own, separating writers cleanly even without explicit stylometric features.
Code →