Ellen Kolesnikova

Decatur High School class of 2026

Medicine: Safe+Natal | Ellen Kolesnikova

Medicine: Safe+Natal

I am working with safe+natal (joint project with Emory University and Wuqu’ Kawoq/Maya Health Alliance) on improving pregnancy healthcare in rural Guatemala through ML-based diagnostics and triage. More specifically, the goal of this project is to provide a simple system that enables rural midwives from underserved communities to conduct basic pregnancy monitoring and identify cases that need to be referred to hospitals. We distributed $200 health monitor kits (including smartphone and very basic devices such as blood pressure monitors and 1-dimensional doppler ultrasounds) to local midwives. Indigenous Maya communities do not commonly use written language, so we automatically transcribe and interpret readings from the health devices to enable timely healthcare decisions. Safe+Natal is tremendously successful: integrated with a Mayan hospital, it reduced maternal mortality from one of the highest in the Western Hemisphere to nearly zero!

My role in this larger project is to develop and test image recognition software that transcribes readings from an image of a blood pressure monitor (taken on the smartphones given to midwives). The challenge in this is to create an accurate system that works with all types of lighting and backgrounds.

Paper linked here.

Crypto: Private Repair of User-Flagged Failures in Text-to-Image Services

This project focuses on using cryptographic tools (namely secure multiparty computation) to securely and privately remove bias from generative AI models. Consider the setting where a private pre-trained model may be biased, e.g., it only generates images of men (not women) in leadership positions. The goal is to use an external party’s advice (which is also private) to fine-tune the model to eliminate the specific bias. We achieve this by computing on encrypted data.

Work done as a part of an internship with Prof. Xiao Wang (Northwestern University); the resulting paper was submitted to NeurIPS 2025.

Digital Humanities: Identifying the author of a Medieval text

The goal of this work (in collaboration with Prof. Juliana Viezure (Georgia Tech) and Jacob Young) is to identify the author of a medieval text - the Passio Et Miracula Sancti Eadwardi Regis Et Martyris. Our work proposes, with compelling evidence, an author for this text, disproving a previous authorship attribution. The new attribution contributes to a better understanding of religious life in England in the period immediately following the Norman Conquest.

Our approach to authorship identification is three-pronged: through (1) traditional literary and historical analysis of the document, (2) standard stylometric analysis based on word frequency, and (3) machine learning. I am solely responsible for the machine learning part of this project; I trained an accurate model to distinguish the author (among four possible choices) of the examined Latin text. When running the Passio through my model, it classified the text with high confidence, corroborating and strengthening my co-authors’ traditional and stylometric analyses.

The corresponding paper is completed and will be submitted to a humanities journal.

Misc. individual projects