Research
Below are the projects I have worked on so far, both published and ongoing, starting from the most recent one. Also on Google Scholar and ORCID.
Equal contribution denoted by *.
Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning
Pietro Carlotti*, Nevena Gligić*, Arya Farahi. STAI-X, 2026. Best Paper Award
Research Question. How to make an AI system more reliable under distributional shifts, when the data it sees comes from a different distribution than the data it was trained on?
Approach. Introduced DIP-EDL, an EDL parametrization that decouples class prediction from the magnitude of uncertainty by separately estimating the conditional label distribution and the marginal covariate density. This separation preserves evidence in high-density regions while shrinking predictions toward a uniform prior for OOD data.
Applications. Fraud detection, medical diagnosis, autonomous driving, and disaster response management — domains where a confidently wrong answer is far more costly than an abstention.
Methods. Normalizing flows (MAF), evidential deep learning.
Robust-by-Design Distributional Learning from Contaminated Samples
Nevena Gligić, Arya Farahi. Under review, 2026.
Paper and code coming soon. · Video
Research Question. How to achieve a distributional objective when the training sample is contaminated?
Approach. CC-MMD (contamination-corrected maximum mean discrepancy) is a design-based estimator that combines cheap, noisy proxy scores over the full sample with audited residual corrections over a small verified subset, to recover the oracle MMD that would have been measured on clean data.
Applications. Galaxy population inference with star contamination, LLM-assisted filtering of political content, and species-distribution inference from environmental DNA.
Methods. Normalizing flows (IAF), VAE, GAN, MLP.
Unsupervised Bayesian Probabilistic Signal Detection in Noisy Environments
Nevena Gligić, Arya Farahi. In preparation.
Paper and code coming soon. · Poster
Research Question. How to detect a signal in a low signal-to-noise ratio (SNR) regime without labels, and without assumptions on the signal distribution?
Approach. BPSD models the data and noise distributions explicitly with normalizing flows and uses Bayesian inference to assign a detection probability to each candidate. Because the signal distribution is never assumed, the framework transfers across domains where the noise is well characterised but the target is not.
Applications. Gravitational-wave detection, galaxy detection, medical imaging, radar systems, and other applications where target signal is not observed clean.
Methods. Normalizing flows (MAF), CNN, GMM, autoencoders.
AI-Driven Utility Monitoring and Anomaly Detection for City-Operated Buildings
Nevena Gligić, Vineet Burugu, Arya Farahi, Matt Kammer-Kerwick. In preparation. In collaboration with the City of Austin.
Paper coming soon.
Research Question. How to detect extreme energy consumption across a portfolio of city-operated buildings in real time?
Approach. A forecasting and anomaly-detection pipeline built on a time-series foundation model, paired with an interactive dashboard that translates each flagged event into estimated cost and CO₂ impact. This allows the people acting on the alerts to triage them by consequence rather than by anomaly score.
Applications. Improved operational efficiency, reduced waste, lower emissions, and lower costs.
Methods. Chronos Bolt time-series foundation model, transformers, ARIMA baselines.
The dashboard is not publicly accessible.