Shreyosi Endow

I earned my PhD in Computer Engineering at the University of Texas at Arlington, specializing in Human-Computer Interaction under the advisement of Dr. Cesar Torres. My research expertise spans natural language processing (NLP), crowdsourcing, generative AI (VLMs and LLMs), and design research. My dissertation contributed automated and scalable techniques for extracting tacit knowledge from instructional content available on the web to better support tutorial search, generation, and navigation.

I currently work as an Applied Scientist contractor on Adobe Firefly’s Scientific Evaluation team, where I co-lead evaluations of generative imaging models and own a scalable data analysis pipeline for grounding evaluation strategy in real-world user behavior. My work informs post-training strategy, model release and acquisition decisions, and PMs’ product requirements.

Throughout my PhD, I participated in several research internships at Adobe Research and Fujitsu Research. You can read more about my work in the research highlights below!

research highlights

Evaluating generative models

At Adobe, I evaluate first- and third-party GenAI models, and I care deeply about grounding evaluations in real-world usage. I own a large-scale data pipeline that analyzes how people interact with generative models, which directly informs our evaluation strategy and PMs’ product requirements. I also own text-to-image evaluations, from building evaluation sets to running human annotation studies and developing auto-evaluators that measure model quality. My evaluations inform model release and acquisition decisions.

Measuring subjective qualities with people and models

Much of what makes a GenAI output good, and what drives customer satisfaction, is subjective, from mood and aesthetics to how well it captures the intent behind a prompt. I measure these qualities by combining crowdsourced human judgments with computational metrics and VLM-as-a-judge approaches, and by studying where human and automated judgments agree and where they don’t. My PhD work explored how the crowd can capture many perspectives on the same tacit experience, and at Adobe Research, I investigated how well VLM judgments of mood in design templates align with human judgments. In my evaluation work today, I build VLM judges for prompt adherence and visual variety in generated images.

Knowledge extraction and personalization

For my dissertation, I studied how people learn hands-on skills from online tutorials, where important details are often left out because they’ve become second nature to the tutorial’s author. Using natural language processing (NLP), LLMs, and VLMs, I extracted workflow knowledge from thousands of video tutorials and developed scalable, low-cost pipelines for modeling each learner’s prior expertise, so that tutorials could be personalized to what they already know.

Also see my dissertation: Beyond Proximal-Distal: A Many-to-Many Model for Tacit Knowledge Transfer in Tutorials (2026).

Creative workflows

I have always been deeply invested in how tools shape the way people learn and make things. I study this through surveys, interviews, and observational studies, and I use what I find to inform how tools and evaluations get designed. During my PhD, I looked at how a tutorial’s medium changes how people develop physical skills like centering clay, and how projected cues can shift what artists notice while drawing. At Adobe Firefly, I brought this into evaluation by building a scalable think-aloud workflow for testing early research prototypes with creative professionals, analyzing their transcripts with NLP and LLMs to surface both qualitative findings and measurable trends.

Embedded systems and sensing

Before GenAI and creative workflows, I spent much of my time building physical interfaces that sense and respond to the body, often borrowing techniques from non-technical crafts like embroidery. I studied computer engineering in undergrad, and I’ve done extensive hands-on hardware work, from designing and soldering circuits to programming microcontrollers at the register level and connecting them to sensors over wired and wireless protocols. That work led to Compressables, a toolkit for prototyping wearable haptic interfaces, and Embr, a framework for making textile displays through electronic hand embroidery. At Adobe Research, I implemented the audio and motion sensing behind Project Primrose, an interactive dress showcased at Adobe MAX.