University of Washington

I teach IA and User Experience classes in the undergraduate and graduate programs at the UW Communication Leadership School and the Informatics School at UW. Students are introduced to facets of the product life cycle, focusing on concepting, interaction design, brand and design systems, user personas, interpreting business requirements and other user centered design concepts including agentic AI frameworks.

I am currently leading a project leveraging Agentic Open Source Intelligence to combat misinformation and strengthening the integrity of media and news as a joint effort between UW, Stanford University, and the National Yang Ming Chiao Tung University in Taiwan. Here are a few of the applications I have built that are being leveraged by professors across UW:

Virtual Teacher’s Assistant

I built for my students to bounce ideas off and receive advice on their assignments. Grounded strictly in my own course material, class transcripts, interviews, and portfolio content, it helps students answer assignment questions between classes and refine their framework docs for their own projects. To build it, I used an AWS stack with Bedrock knowledge base services, vector search, and custom data processing, converting recordings into tagged markdown, chunking them, and syncing them via S3. I also explicitly engineered a "walled garden" persona prompt to keep the model strictly on-topic and set its memory behavior to zero to keep database overhead manageable and as an MCP to leverage for other application uses to confirm with.

lecturebot.perfectpixels.com

Agent Behavior Maker

To assist students in building LLMs, I built a tool that creates synthetic customers to test and refine an AI persona's formality, verbosity, and empathy before engineering builds against it. This fits into my broader framework of treating agent behavior as a design material across eight key dimensions: purpose, role archetype, initiative level, capabilities, interaction context, memory behavior, tooling, and boundaries. In class, students use this process to generate the precise steering document handed off to build tools like Claude Code, serving as a core assignment deliverable that defines the personality and experience level of the AI.

To use it, you load a product proposal, user personas, and agent framework, and the tool dynamically generates five core user scenarios based on your inputs. You then pick a scenario to watch a synthetic conversation play out between the user and the bot, allowing you to turn dials like empathy, verbosity, proactivity, and technical depth to re-render the exchange in real time. Once you settle on how those behavioral traits actually read, you export the finalized steering document and drop it into your build pipeline.

ai-persona-builder-seven.vercel.app

Video Annotator

I built a feedback tool of prototype walkthroughs. Building prototypes are more difficult for receiving feedback and collaborating with others on, versus mocks in a tool like Figma. Students screen-record themselves talking through a prototype, but a static video is a dead artifact that makes inline feedback impossible. To fix this, I brought Figma-style commenting to prototype walkthroughs so anyone can mark a comment, question, or edit request at a specific timestamp and reply directly.

In practice, you simply record and upload your screen-capture walkthrough, pause where you want to leave a note, commit comments like "Problem here" or "Too dark," etc., and use the side panel to jump directly to flagged moments. You can then export all these notes into a Word document to share feedback and collaborate seamlessly. The main technical hurdle was handling YouTube videos, where pulling video in worked initially but capturing screenshots failed, requiring us to test several technical approaches before landing on a functional solution of storing the videos on S3 directly versus streaming from an outside service.

annotator.perfectpixels.com