Eight years shipping production systems — healthcare, then Meta.

Architected a dual-transformer system pairing DeBERTa-v3-Large (434M) and RoBERTa-Large (355M) with agreement-based routing; LLM arbitration triggered only when models disagree (~1% of queries), reframing the LLM's task from ternary to binary classification. Avoids the failure mode of confidence-based routing on overconfident model errors. Key finding: architecture matters more than scale, and inter-model disagreement is a better uncertainty signal than softmax confidence. Presented at Johns Hopkins Carey Business School national championship, February 2026.

Independent research investigating whether persona framing like "You are an expert in X" — a norm borrowed from human collaboration contexts — imposes an unnecessary scarcity constraint on LLMs. Core hypothesis: telling an LLM to be a specialist doesn't activate specialized knowledge; it artificially suppresses everything else, potentially narrowing the conceptual search space on generative ideation. Refined task-dependent hypothesis: expertise framing is neutral-to-noise on retrieval-heavy factual tasks, beneficial for structured reasoning, potentially harmful for creative/generative ideation. Measurement pipeline uses Shannon entropy over a 15-domain taxonomy.

Structured decision-making tool: AI personas (optimist, pessimist, analyst, creative, synthesizer) offer distinct perspectives in parallel while the user provides emotional context as the Red Hat. The Blue Hat (synthesizer) integrates all inputs including the user's own. Underlying data model supports custom personas, custom panels, weighted ensembles, and alternate deliberation frameworks beyond Six Hats. Provider-abstracted AI layer supporting both Anthropic Claude and OpenAI GPT-4o.

Open-source Python library for confusion-matrix analysis and ML meta-research auditing. Features include metric reconstruction, probabilistic inference, statistical testing, and cost-sensitive analysis. Designed to catch errors and implausible results in published ML evaluations. Reflects an applied interest in ML reliability and the integrity of published evaluation metrics.

Built and maintained a COVID data warehouse aggregating data from all 46 Prime Healthcare hospitals to support corporate decision-making and government reporting. Developed automated report interfaces and data feeds; surfaced proactive savings opportunities and value analyses for contracts under review. Most of the original architectural work I did pre-pandemic was compressed into a few months when the warehouse became operational under time pressure.