15+
Selected research
Zero-Shot Epidemiology: Deep Learning & Time Series Forecasting
This project pushed beyond straightforward forecasting. The goal was to assess whether a model could remain useful even when disease-specific training data was unavailable. The work required careful model design, benchmarking and statistical validation, not just model fitting.
5
Best Paper
Overview
An award-winning forecasting framework built to generalize across pathogen types without disease-specific training data.
How do you produce robust forecasts when you cannot depend on disease-specific labeled data? The challenge was to design a model that generalized, then prove that its behavior was numerically sound and understandable to stakeholders.
01
Defined a zero-shot forecasting framing using transfer learning and deep learning sequence modeling.
02
Built evaluation loops that compared the framework to multiple baseline models rather than relying on a single headline score.
03
Measured not only point-error metrics such as RMSE and MAE, but also confidence intervals and assumption stability under shifted conditions.
04
Documented limitations, risk factors and where the method remained dependable versus where caution was needed.
01
Created an award-winning research framework with both predictive and analytical depth.
02
Showed strong emphasis on validation, reliability and benchmarking — not just raw model output.
03
Translated advanced technical findings into a technical report suitable for public-health-oriented stakeholders.
Detailed notes
What the work demonstrates.
Problem framing
The work began with a difficult forecasting constraint: the system had to operate without disease-specific training data. That made transferability and statistical robustness central to the solution, not secondary considerations.
Modeling strategy
The forecasting pipeline was designed in PyTorch with transfer-learning concepts that could absorb useful temporal structure from broader patterns while remaining adaptable across pathogen classes.
Validation
Performance was benchmarked against five alternative approaches using RMSE, MAE and interval-based reasoning. Hypothesis testing was also used to assess whether the model remained reliable when distributions shifted.
Why it matters
This project reflects the kind of work where modeling, experimentation design and communication all matter equally. It demonstrates strong alignment with research-heavy AI and data science roles.