← Home
Human-AI interaction, voice interfaces, and what happens when retrieval has to work on a real corpus rather than a benchmark.
Two published, four in preparation
Speak a restaurant review and let an AI help turn it into one you would post. A field study with 14 people and 157 reviews: willingness to share rose 31%, and confidence in writing a good review rose 88%.
doi.org/10.1145/3701571.3701593 ↗ (opens in a new tab)What students and staff would need before trusting a social robot on campus, from 59 people who met a Pepper robot first. Knowing what the robot does, and who is behind it, came first.
doi.org/10.1145/3764687.3764717 ↗ (opens in a new tab)Finding the moments a public conversation turns, and giving that evidence enough structure that something else can act on it. The classifier behind it also runs inside NeuroClima.
Preprint on arXiv ↗ (opens in a new tab)Reading the explanations behind an LLM judge’s faithfulness scores for 1,197 answers from NeuroClima, and finding what a single number leaves out: about one in seven pieces of evidence the judge relied on was text a model had written, not the source.
Asking questions of the research literature and getting answers grounded in a knowledge graph assembled from several sources, rather than a single index.
Generating intrusion detection signatures from logs automatically with large language models, while keeping each signature explainable and interpretable.