I build AI products that turn messy workflows into usable tools.
My current work sits at the intersection of AI agents, human-computer interaction, creator tooling, and local-first productivity systems. I care about products that are not just technically interesting, but actually usable in real work: clear workflows, low friction, reliable feedback, and strong human control.
- AI agent tools for coding, research, automation, and personal workflows
- Multi-agent and multi-model workspaces for switching between Codex, Claude, OpenCode, Gemini, and local models
- Voice-first coding and dictation tools for faster interaction with AI systems
- AI content production pipelines for video, comics, short-form content, and knowledge capture
- HCI-informed product design: reducing cognitive load, making AI behavior visible, and turning prototypes into usable products
A CLI AI model profile switcher for Codex, Claude, OpenCode, Gemini, and local models.
It includes model recipes, shell helpers, API presets, shared memory, secret auditing, and portable state management.
Why it matters: AI builders often work across several model providers and agent tools. This project explores how to make that switching reliable, repeatable, and less painful.
A Windows voice input app for coding dictation, supporting local SenseVoice and online Qwen-ASR Realtime.
Why it matters: Coding with AI increasingly feels conversational. Voice input can reduce interaction cost when writing prompts, explaining changes, or managing long agent sessions.
A local desktop workflow for judging whether Bilibili learning videos are worth watching, collecting, and saving into Obsidian.
It combines a browser extension, local Flask backend, native desktop interface, subtitle capture, AI-based learning value analysis, and Obsidian integration.
Why it matters: Learning from video is noisy. This project explores how AI can help filter, evaluate, and preserve high-value learning material in a local-first knowledge workflow.
An AI-assisted comic and video creation platform built around script generation, storyboard generation, character consistency, and creative evaluation.
Why it matters: AI content tools need structured workflows. This project explores how script, storyboard, visual assets, and evaluation can become one production loop.
A browser tool for extracting subtitles from Bilibili videos, previewing them, exporting Markdown, downloading subtitle files, and writing notes into Obsidian through Local REST API.
Why it matters: Knowledge capture should be fast and local-first. This project connects video learning with personal knowledge management.
A turn-based civic strategy simulation game using LLM-driven citizens and value conflicts.
Why it matters: It combines HCI, simulation, sustainability, and AI explanation to explore how games can help people reason about complex social trade-offs.
- AI Agent UX: task state, progress visibility, error recovery, and human override
- AI workflow products: from prompt experiments to repeatable systems
- Local-first tools: privacy, ownership, portable data, and personal automation
- Creator tools: AI video, comics, short-form content, and media workflows
- Developer tools: CLI design, desktop apps, automation, and multi-model environments
Python TypeScript React Node.js Rust Tauri Shell
LLM APIs Multi-Agent Systems Voice AI Local-first Apps Workflow Automation
- MSc Human-Computer Interaction, Newcastle University
- BSc Computer Science, Heriot-Watt University
- Product-minded builder with a computer science and HCI background
I like working on the layer between model capability and real user behavior: where AI systems need structure, feedback, interaction design, and product judgment before they become useful.
- Email: albertexert@gmail.com
- GitHub: YuxuanSun123
- Portfolio: yuxuansun.netlify.app