CS grad. Interested in where systems break and why.
- building evals and agent reliability systems — making AI behavior predictable before it ships
- researching how small transformers develop (and fail to generalize) reasoning procedures
- exploring the gap between in-distribution performance and real-world reliability
sovereign-rag-ratchet Self-improving RAG pipeline on Cohere's stack (Embed + Rerank + Command R+) with a guardrail that catches metric-gaming before it ships. Built for air-gapped / sovereign deployments. Model proposes, code disposes.
watt-bench Open-source benchmark for power-constrained LLM inference scheduling — tokens-per-watt as a first-class metric, rack-level power budgets as hard constraints. Thermal cascade demo: greedy_latency triggers 1,000+ throttle events vs greedy_power's zero on real Azure traces.
cot-scratchpad-research Trained three 0.79M-parameter transformers under Direct, CoT, and Scratchpad paradigms. The Scratchpad model spontaneously developed carry notation — then truncated its reasoning at exactly 2 digit steps on 3-digit inputs. Evidence of learned procedure, not generalizable algorithm.
CS @ Dickinson College '26