- Built an end-to-end pharmacovigilance pipeline over massive unstructured social-media corpora, using transformer models and weak supervision to surface drug-safety signal that structured reporting misses.
- Fine-tuned transformer architectures and hand-rolled programmatic labeling workflows before ChatGPT made either one a standard tool.
- Published and presented the methods through peer review and national conference talks.
The doctorate was natural language processing, done in the window just before
the generative wave broke. Transformers were still a research instrument then,
not an API. That meant learning the machinery from the inside: tokenization,
fine-tuning, and the specific ways a model trained on noisy human text goes
wrong.
The timing mattered more than it looked at the time. Production LLM and RAG
systems now run into the same walls the dissertation ran into daily: evaluation
without ground truth, drift, the distance between a clean benchmark and a real
corpus. Tooling changed completely. The hard parts did not.
- Concentration in infectious disease epidemiology: outbreak investigation, transmission dynamics, and the study designs that let you argue causality from observational data.
- Emergency preparedness and disaster response, which is decision-making for low-frequency, high-consequence events under badly incomplete information.
Epidemiology trains one reflex above all others: ask what would have to be true
for this association to be causal, then go looking for every reason it is not.
Applied to model performance, that reflex separates reporting a metric from
understanding one. A headline number is a hypothesis; it is not a finding.
Preparedness training left a second habit that transfers cleanly to running AI
systems. Plan for the tail. Rare, severe failures are the ones worth real design
effort, and they are exactly the ones an average-case evaluation will never
show you.
- Undergraduate research on bat hibernation: field data collection under conditions where the measurement is the hard part, not the analysis.
Where the empirical habits started. Studying hibernating bats is field science at
its least forgiving, with small samples, hostile conditions, and animals that
decline to cooperate with the protocol. Two lessons stuck. Data is something you
go and get; and most of the uncertainty in a result was introduced long before
anything reached a spreadsheet.