Research

• Uncertainty quantification and safety in LLM-based health-coaching agents2026

  At JSALT 2026 I worked on the uncertainty and safety side of building a health-coaching agent. I designed an evaluation set by injecting examples and multiple-choice questions into real coaching dialogues, worked on the safety system, and developed an uncertainty method, based on disagreement across several models, that tells the agent when it should ask a clarifying question. In parallel I worked on the dialogue-level methods below.

  papertalk
• Dialogue-level uncertainty quantification for LLM agents2026 –

  Building uncertainty quantification methods for LLM-based conversations that take into account the effect of the user, with evaluations focused on the actions uncertainty should trigger, such as asking the user or handing off. Work in progress.

• CASPER2023 – 2025

  An English spontaneous speech corpus with 208 speakers. I worked on it for three years and helped shape how the corpus was built and released. I built its database and the processing pipeline that turned raw recordings into the released corpus, and wrote major parts of the paper.

  paperdataset
• Text-only speaker attribution for multi-talker ASR2025

  At JSALT 2025 I was on the End-to-end Multi-channel Multi-talker ASR (EMMA) team. I worked on finding out who said what in a conversation from the transcript alone, with no audio. The goal was to give speaker labels to speech recognition systems that output none, including one the team developed, and to probe how well LLMs can do speaker attribution on their own.

  talk
• Aruz Scanner for Ottoman Divan poetry2026

  This is a passion project I started after realizing that the automatic tools for aruz scanning are limited and unreliable. The resulting open-source tool, based on an HMM over reading lattices, finds the meter of a Turkish poem written in aruz and gives its scansion, along with the flaws it finds.

  toolcode