Postdoctoral Research Fellow · University of British Columbia · Vancouver, BC

Making AI systems
accountable to the
people they overlook

I work on responsible AI across three fronts: fairness in recommender systems, bias in large language models, and privacy in NLP. My PhD built the methods. My postdoc applies them where the stakes are highest, in global health.

Portrait of Shera Potka
01 / Fairness

Recommenders decide who gets seen

Link recommendation shapes who finds whom in a social network. Left alone, it compounds advantage: popular nodes get recommended, become more popular, and get recommended again. Structural minorities disappear into the tail.

  • MinWalk raises the visibility of structural minorities while holding utility loss to a minimum.
  • Reduces popularity bias without discarding the signal that makes recommendation useful.
  • Best Paper, MEDES 2024.

MinWalk — structural minority visibility in link recommendation. Highlighted nodes are minorities surfaced by the walk.

SC-WEAT — gender and race association in embedding space, measured across models.

02 / Bias

Language models carry the bias of what they read

Bias in an LLM is not a rumour to be argued about. It is a quantity that can be measured, in embedding geometry and in what the model actually recommends to people. I measure both.

  • Gender and race association in modern embeddings, via SC-WEAT and clustering.
  • Bias in consumer product recommendations made by LLMs, where it reaches users directly.
  • Published at I-SPAN 2025 and AINA 2025.
03 / Privacy

Text can be private and still mean something

Differential privacy protects people by adding noise. Add it carelessly and the text stops saying anything. The real problem is not privacy or utility, it is the exchange rate between them.

  • CluSanT — text sanitization under Metric Local Differential Privacy.
  • Clustering plus embeddings hold semantic coherence at a given privacy budget.
  • Published at NAACL 2025.
Two heatmaps comparing CluSanT against CusText at privacy budgets epsilon = 1 and epsilon = 8, showing semantic similarity gains

CluSanT vs CusText — semantic similarity gain at ε = 1 and ε = 8. Warmer means more meaning retained at the same privacy budget.

patient provider 19 topics · 2 clinical contexts

WelTel — two-way SMS between patients and providers, Rwanda and Canada.

04 / Applied

The same questions, where the stakes are real

At UBC I work with the Digital Global Health & AI Group on real patient–provider messaging. Bias and privacy stop being abstractions when the data is a person texting a nurse in Kinyarwanda.

  • Multilabel topic classification across 19 topics in two clinical contexts, Rwanda and Canada.
  • Benchmarked six machine translation systems on real Kinyarwanda clinical dialogue. None is safe unsupervised.
  • Bounded claims by design: human-in-the-loop support, never autonomous triage.
05 / Building

Research that has to run

Methods that never leave a paper are hard to trust. These are the systems where the ideas meet users, latency, and consequences.

Visual analytics

ConVisScope

A research workbench turning WelTel's two-way SMS stream into population-level intelligence for Rwanda's digital health ecosystem. Cohort filtering, label review and correction, and reproducible export, generating the validated correction stream that trains the next classifier.

UBC mHealth Lab · deployed on AWS · 2026–present
Clinical search

NoSpin-EBM

Evidence-based medicine search for clinicians, built around fail-closed provenance: a field renders only when it is anchored to a retrieved record. GRADE-rated evidence cards with verified DOIs, and a Skeptic's Corner that flags bias in the underlying evidence.

FastAPI · React 19 · Crossref/PubMed/openFDA · 2026–present
Clinical prototype

PageLink

An auditable replacement for the hospital pager: paging and acknowledgement, on-call scheduling, in-app consult chat and structured clinical coding, designed under Canadian privacy constraints. Built as the artifact that scopes a production MVP.

React prototype · with Dr. R. Lester · 2026
06 / Teaching

Training the people who build the next systems

I have designed and taught a core graduate course across three offerings, and supported undergraduate cohorts of up to 180.

  • Postdoctoral Research Fellow May 2026 – Present
    Digital Global Health & AI Group, University of British Columbia

    AI and NLP research on patient–provider mobile-health data. Topic classification and clinical-language models supporting real digital health programmes, plus AI implementation strategy across the group's mHealth platforms.

  • Instructor — Data Models & Algorithms (CSC 501) Spring 2025 – Spring 2026
    University of Victoria

    Designed and taught this core graduate course across three offerings (≈50 students each), emphasizing active learning and algorithmic problem-solving.

  • Lead Teaching Assistant — Data Mining (SENG 474) Spring – Fall 2024
    University of Victoria

    Coordinated teaching support across three terms, class sizes 50 to 180. Labs in classification, clustering, and pattern mining.

  • Researcher Nov 2022 – Nov 2023
    Cologne Center for eHumanities, University of Cologne

    Workflows for large-scale text and metadata analysis in digital humanities. Advanced Project RACIR through tooling and cross-disciplinary research.

  • Research Fellow Oct 2022 – Mar 2023
    Electronic Textual Cultures Lab, University of Victoria

    Visualizations and computational tools for digital scholarship. Directed the HSS Commons initiative, shaping platform design and implementation.

07 / Record

Publications & manuscripts

Under review

  • Selective Overconfidence: How Large Language Models Underserve Social Welfare Law Across Jurisdictions

    Shera Potka, Alex Thomo, Wulf Loh, Paula Helm

    AIES 2026 · AAAI/ACM Conference on AI, Ethics, and Society

    Under review
  • People, Processes, Platforms: A Coding Framework and Comparative Benchmark for Global AI Governance

    Shera Potka, Jens Weber

    AIES 2026 · AAAI/ACM Conference on AI, Ethics, and Society

In preparation

  • Multilabel Topic Classification of Patient Messages Across Two Clinical Contexts

    Riyaz Ur Rehman, Andy Liu, Hyeju Jang, Hassan Mugabo, Shera Potka, Richard T. Lester

    Targeted at npj Digital Medicine

    2026
  • Evaluating Off-the-Shelf Machine Translation for Kinyarwanda-to-English Clinical Dialogue

    Hassan Mugabo, Riyaz Ur Rehman, Andy Liu, Matthew Manson, Justin Tuyisenge, Muhammed Semakula, Shera Potka, Richard T. Lester

    Targeted at The Lancet Digital Health

    2026

Peer-reviewed

  • CluSanT: Differentially Private and Semantically Coherent Text Sanitization

    Ahmed Musa Awon, Yun Lu, Shera Potka, Alex Thomo

    NAACL 2025

  • Gender and Race Bias in Consumer Product Recommendations by Large Language Models

    Ke Xu, Shera Potka, Alex Thomo

    AINA 2025 · Advanced Information Networking and Applications

    2025 Paper →
  • Word Embedding Bias in Large Language Models

    Poomrapee Chuthamsatid, Shera Potka, Alex Thomo

    I-SPAN 2025 · Pervasive Systems, Algorithms, and Networks

  • Enhancing Structural Minority Visibility in Link Recommendations

    Shera Potka, Isla Li, Jason Kepler, Alex Thomo

    MEDES 2024 · Management of Digital EcoSystems · Best Paper Award

    2024 Paper →
  • Community Structure and Coherence in Digital Humanities Works

    Shera Potka, Alex Thomo

    IISA 2023 · Information, Intelligence, Systems & Applications · Best Paper Award

    2023 IEEE →
08 / Contact

Let's talk

Open to collaborations, research partnerships, and teaching. If your work touches fairness, bias, privacy, or health data, I would like to hear about it.

Vancouver, British Columbia
Division of Infectious Diseases, UBC

github.com/sherapotka
linkedin.com/in/shera-potka