Large language models are quietly changing two things I care about: how science gets done, and how people learn. A lot of the hard work in both is repetitive, judgment-heavy, and badly bottlenecked by human time - exactly the kind of work that well-designed AI can take on, if you build it carefully. I work on both ends of that: tools that speed up research, and tools that make learning more personal and accessible.
PorAI is an educational AI initiative built around a simple idea: every learner deserves the kind of attention a great personal tutor provides. It brings together students, teachers, parents, and institutions on one platform, using AI to make instruction adaptive - meeting students where they are, surfacing where they're stuck, and giving teachers and parents a clearer picture of how learning is actually going. Having spent years in olympiad mentoring and quantum-computing outreach, I see PorAI as the same mission at a larger scale: lowering the barrier between a curious person and a good explanation.
Egent (Ting, Saad, Liu & Shen, 2025) is an autonomous agent for measuring equivalent widths of stellar absorption lines - one of the most labor-intensive steps in detailed chemical abundance work. It combines classical multi-Voigt profile fitting with an LLM that visually inspects each fit, catches failures, and iteratively refines the result, all directly on raw flux spectra. Benchmarked against expert measurements across 84 stars and more than 18,000 lines, it reproduces human-level results in days rather than months. The code is open source, with both a web interface and an offline version that runs on a laptop, so it can scale to large spectroscopic surveys.
Saad & Ting (2026) takes the agentic idea a step further: instead of one bespoke pipeline, it packages spectroscopic analysis as a set of small, reusable tools that an LLM agent can call and combine on its own. As a first application, the agent searches for unresolved binary stars hidden in survey spectra, flagging more than 41,000 candidate double-lined spectroscopic binaries in APOGEE DR19. The same toolset is designed to carry over to other survey-scale measurement tasks.
Egent is one piece of a fast-emerging shift toward agentic science, where AI systems don't just predict numbers but carry out multi-step research tasks with judgment and self-correction. I'm interested in pushing that further across my own work - automating the tedious parts of spectroscopy, abundance analysis, and survey-scale data reduction - so that researchers spend more time on the questions and less on the machinery.