I'm Aishwarya — a doctoral researcher, a former full-stack engineer, and this year's lead TA for CU Boulder's CS department. I build things, study why they work, and teach other people to do both.
We have spent decades teaching machines to optimize: clicks, purchases, watch time, accuracy, engagement, efficiency. These metrics have made our systems remarkably powerful, but they are proxies for human goals, not human values. If we want AI to reflect humanity, someone has to decide what it's values are, and make sure that answer comes from the people and communities it serves. That's the work I care about.
My dissertation asks what a news recommender looks like if it optimizes for something other than clicks.
No infrastructure existed for longitudinal recommender research with actual news readers.
More diverse recommendations usually mean fewer clicks — a tax that keeps diverse systems from shipping.
Journalists' values are absent from recommender design — no method existed to encode them as inputs.
Fairness metrics are usually designed by developers, not the providers a platform actually affects.
Satwani, A. Aligning Algorithms with Axiology: Operationalizing Journalistic Values in News Recommender Systems.RecSys 2026 Doctoral Symposium — Accepted
Satwani, A., & Burke, R. Personalizing Diversity in News Recommendation: A User Study.RecSys FRAME Workshop 2026 — Accepted
Higley, K., Burke, R., Ekstrand, M., Kluver, D., Knijnenburg, B., Malthouse, E., Hasan, M., Iqbal, M., Parajuli, S., Satwani, A., Sristy, Z., Suna, R., & Konstan, J. Designing HCAI Field Experiment Research Platforms: The POPROX News Experience.International Journal of Human–Computer Interaction — Under Review
Smith, J. J., Satwani, A., Burke, R., & Fiesler, C. Recommend me? Designing fairness metrics with providers.ACM Conference on Fairness, Accountability, and Transparency (FAccT '24), 830–839
Alaqabawy, N., Satwani, A., Voida, A., & Burke, R. "It's just a robot that looks at numbers": Restoring journalistic voice in news recommendation.INRA @ RecSys 2023 — ceur-ws.org/Vol-3561/paper1.pdf
Alaqabawy, N., Satwani, A., & Burke, R. The Journalistic Algorithm: A news recommender platform for public service journalism.Broadcast Education Association
Satwani, A., Rao, M., & Nagaraju, N. Data Mining Techniques for Heart Disease Prediction.International Conference on New Trends in Engineering and Technology (IEEE-supported), Chennai
Personalized diversity (MMR-P) holds engagement steady while surfacing far more variety than generalized diversity does. Here's a tiny working version of that idea, pointed at my own work instead of a news feed: drag the dial and watch it re-rank between "most relevant to my PhD" and "most diverse across everything I do."
The flagship finding: personalized MMR-P holds engagement steady while maximizing diversity, where generalized MMR just tanks clicks.
Co-built an open experimentation platform with 5 universities and the Associated Press.
Providers experience unfairness as underexposure, overexposure, and lost agency — none captured by standard fairness metrics.
Took the annual increment process from 3 months of manual Excel work to a live web app — end to end, as sole engineer at first.
Mentoring student teams through the full software lifecycle — not just the code, the judgment calls around it.
The non-algorithmic side of paying close attention — painting when the ranking metrics can wait.
Before the PhD, and running alongside it, I've been a working engineer — three years at Walmart Global Tech taking a solo intern prototype to a product used by thousands, plus a forecasting engine at Hitachi Energy. The research asks what recommenders should value; this is proof I can ship the thing.
Smaller builds outside of work and the dissertation — recommenders, optimization, and NLP. Code linked where public.
DKN, NRMS, and LSTUR on the MIND dataset with MMR-based diversity re-ranking.
View on GitHubCross-country route planning with a Genetic Algorithm, handling travel-limit and stop constraints.
View on GitHubLDA topic modeling over 1.2M customer reviews, distilled into 14 features per category with feature-level sentiment.
View on GitHubRoBERTa-based binary classifier for Patronizing and Condescending Language, built for SemEval 2022 Task 4.
View on GitHubUser-user collaborative filtering on MovieLens with Euclidean similarity and Laplace smoothing, evaluated on RMSE and NDCG.
View on GitHubThis year I'm lead TA for CU Boulder's software engineering capstone — mentoring student teams through the same full lifecycle I write about in my research: not just the code, the judgment calls around it.
Oil paintings, charcoal sketches, pencil and lead-pencil work — the non-algorithmic side of paying close attention. A running gallery, added to over the years.
Open to research collaborations, engineering roles, and conversations with people who think recommender systems should serve people first.