Thinking
Papers in one column, essays in the other.
Everything I've written since I was nineteen has been one long question: what is the difference between technology that connects people and technology that imitates connection? — the throughline
In flight and out the door.
Four years across UC Santa Cruz and Google Research — first on the ethics of algorithmic media, then on whether large language models can be safely useful inside health and well-being contexts. The throughline has been the same the whole time.
- Draft · planning CHI 2025 submission · with Shalini Verma
Designing for Internal Authority: AI Coaching Systems as a Counter to the Nodding Mirror
AI systems trained on user-approval signals produce warm, validating responses that provide short-term relief while eroding the user's capacity for self-directed emotional processing. This paper introduces internal authority as a primary design criterion for AI coaching systems and presents a five-move design framework grounded in evidence from a working coaching cohort (N=18, Spring 2026). Drafting now; planning a CHI 2025 submission.
↳ Read the draft on GitHub Code, benchmark, and fine-tuned model on /build
- Frontiers in Psychiatry · Submitted · April 2026
Addiction as a Disorder of Connection: An Interdisciplinary Synthesis of Neuroscience, Trauma, and Social Capital
The neurobiological disease model of addiction fails to explain three empirical anomalies: high rates of natural recovery, the graded dose-response between relational trauma and addiction risk, and the dramatic effect of social environment on compulsive use. This review synthesises five evidence streams to argue that addiction is fundamentally a disorder of connection.
- AAMAS 2024 · Published
Foodlens: computer vision for food recognition
Co-authored work from the Google Bangalore food-app vision team. Studies how vision-model errors translate into usable nutrition signals in low-resource consumer apps.
- CHI 2025 · Submitted
Coaching LLMs and AI ethics in health contexts
First-author submission from a year-long Google Research project on whether LLMs can be designed to coach users toward well-being without creating dependency. Mixed-methods study with the Fitbit team.
- CAVEAT (UCSC) · Tool
Twitter Analysis Tool (TAT)
The research instrument I built at CAVEAT to surface algorithmic bias and echo-chamber dynamics on Twitter. Used in the center's first round of empirical studies.
The benchmarks, datasets, and code behind these papers live on /build.
The thing I can't say in a paper.
Long-form, slowly. Research has a voice. Coaching has another voice. Neither of them is mine when I'm being honest. The essays are where I try the third one.
New essays land on Substack first and sync here on the next build. Nothing showing right now — check back, or read directly on Substack.