I study how learning systems work—and build systems to test what they can reliably do.
I'm completing a B.S. in Applied Mathematics & Statistics at Bryant University (December 2026), and joining Opportunity Insights at Harvard University as an Associate Data Scientist in January 2027. My interests span reinforcement learning, mechanistic interpretability, and reliable AI systems.
Portfolio · Interactive ML demos · LinkedIn · Email
- FastCombo — Deep reinforcement learning for price discovery in partially observed combinatorial auctions. Preprint forthcoming.
- Black Box to Whom? — Investigating what constitutes evidence of mechanistic understanding of language models, through causal interventions, held-out predictions, generalization, and coverage.
- SafeWalk — An open-source iOS and backend system for comparing shortest and risk-weighted walking routes using public incident data. Combines spatial modeling, graph routing, and multicity evaluation with strict temporal holdout.
- Ask Tupper — A campus assistant combining Qwen3-32B QLoRA fine-tuning, hybrid retrieval, and layered prompt-injection defenses.
- Neural networks from scratch — Automatic differentiation, multilayer perceptrons, and character-level language models built from first principles.
I contribute to the tools used to train, evaluate, and study learning systems.
| Project | Contribution |
|---|---|
| TorchRL | Implemented shared MLP output bias and strengthened numerical tests for RL and LLM losses. |
| Mava | Unified advantage estimation in MAT and Sable, alongside fixes to JAX compatibility and training metrics. |
| Jumanji | Added a public observation-from-state API across environments and wrappers. |
| TorchGeo | Corrected self-supervised augmentations for standardized inputs. |
| UniRL | Fixed Hugging Face checkpoint resolution for meta-initialization. |
Previously, I worked on model governance and LLM-assisted underwriting workflows at MAPFRE Insurance, and forecasting, computer vision, and semantic search at Rhode Island Novelty.