GoalPref-Bench
Evaluation harnessA benchmark that tests whether nine language models stick to what a user actually wants or just agree with them. It turns a common worry about sycophancy into one number you can compare across models.
Open to faculty, research & data science roles · Boston, MA
Sean W. Kelley
Human–AI interaction · behavioral science · mental health
I measure how AI systems behave, and build tools that act on what I find.
Postdoctoral researcher at Northeastern's Network Science Institute and co-founder of Myndgard. My research measures how language models behave — things like sycophancy, epistemic independence, and causal reasoning. At Myndgard, I build mental-health tools that clinicians use in their work.
About
I work across human–AI interaction and behavioral science. My PhD is in psychology, focused on computational psychiatry — building machine-learning and network models of mental health from real longitudinal data. That work led to first-author papers in Nature Communications, PNAS, and NPJ Digital Medicine.
Now my research turns vague claims about how language models behave into numbers you can compare across models. I also co-founded Myndgard, where I build AI safety and mental-health tools that universities use with their counseling services.
My work splits in two: research on how AI behaves, and products that put it to use where it matters most — mostly mental health.
- Now
- Postdoctoral FellowNortheastern · Network Science Institute
- Founder
- Co-founder & Technical LeadMyndgard
- Trained
- PhD, Computational PsychiatryTrinity College Dublin
- Funding raised
- $790K+ as PIacross 4 competitive grants
Research
Measuring how language models behave.
Causal-Coherence Probing
Framework + studyAsks a model to forecast something and explain its reasoning as a causal diagram. It then works out which factors matter most and challenges them one at a time: does the model change its mind more when you push on an important factor than a minor one? Run across seven models and hundreds of questions, with reliability and placebo checks.
Personalization & Epistemic Independence
Controlled experimentsExperiments on what happens when you personalize a model to a user. Personalizing makes it warmer and more emotionally in tune, but whether it also makes the model more agreeable depends on the role it's playing.
Personalized AI for Creative Work
Human–AI collaborationA study with hundreds of people testing whether a personalized AI helps with creative work. It did: people and the AI built on each other's ideas across a back-and-forth, instead of the AI just doing the work. I ran it end to end, from the question to the analysis.
Products
Built at Myndgard, shipped and in development.
Discover
DeployedA 12-day, 38-module mental-health course. An AI pipeline writes the content for a range of student backgrounds and checks its own quality. 90% of students improved on mental-health measures.
Share
DeployedMonitors each student's mental health during the wait for care and flags those who are deteriorating to the clinic manager — so students getting worse are seen sooner, not triaged by who booked first. It also gives the therapist a pre-session summary. In production with university counseling services; 15–20% shorter initial sessions and fewer total sessions per student.
Guardian
In developmentWatches a teenager's conversations with AI companions for signs of harm, and runs on the device itself so the conversations never leave it.
Discover Family
In developmentA 10-module course that teaches teens (13+) to spot the ways AI companions can go wrong — flattery, growing attachment, blurred personas, being pulled away from real relationships, and bad responses in a crisis. Each module has scenarios to work through and an AI coach to talk them over with.
- 2026Personalization Increases Affective Alignment but Has Role-Dependent Effects on Epistemic Independence in LLMs.
arXiv:2603.00024 · under review
- 2025Personalized AI Scaffolds Synergistic Multi-Turn Collaboration in Creative Work.
arXiv:2510.27681 · under review
- 2023
- 2022Using language in social media posts to study the network dynamics of depression longitudinally.
Nature Communications, 13(1)
- 2022Machine learning of language use on Twitter reveals weak and non-specific predictions.
NPJ Digital Medicine, 5(1)
Side projects
Quick experiments on questions I'm curious about.
LLM evaluation · clinical intake · synthetic patients
What can an LLM intake agent surface that the patient never volunteers?
Reasoning models surfaced far more of the hidden symptoms: 56% for Kimi K2.6, against 14–18% for standard chat models.
Experience
- Sep 2024 — PresentPostdoctoral Research Fellow, Human–AI InteractionNortheastern University · Network Science Institute
- Dec 2022 — PresentCo-founder & Technical LeadMyndgard
- Oct 2018 — Dec 2022PhD Researcher, Computational PsychiatryTrinity College Dublin
Education
- 2023PhD, PsychologyTrinity College Dublin
- 2018MSc, Medical NeurosciencesCharité — Universitätsmedizin Berlin
- 2016BS, Neuroscience & Behavioral Biology (Highest Honors)Emory University
Grants & awards
- Enterprise Ireland Commercialisation Fund$629K · PI
- Enterprise Ireland Proof of Concept$143K · PI
- Enterprise Ireland Feasibility Study$16K · PI
- Schmidt Sciences — AI at Work$10K · PI
- Trinity College Dublin Provost's PhD Award$132K
- NeuroCure Cluster of Excellence Scholarship$14K
- EIT Health Wild Card1 of 8 EU startups
- Dogpatch Labs Founders ProgrammeResident