CorpusStudio
A writing environment that shows how others structured the same kind of text, while you write your own. It surfaces sentence patterns and section conventions from a corpus of prior papers, right at the cursor.
Hi, I'm Hai.
These days I'm building AI systems and thinking a lot about where human-AI interaction goes next.
Over the last few years I've worked on AI for writing, creativity, enterprise workflows, and real-time assistance. Along the way I earned a PhD in human-computer interaction, led research projects at Autodesk, Meta Reality Labs, and Harvard, and later co-founded two startups, including one backed by Y Combinator.
This site is where I collect projects, notes, and half-formed ideas as I try to make sense of it all.
Index
- 2026Something new, not public yetnow
- 2026runMachine, AI agents for B2B customer supportco-founder
- 2025CorpusStudio: writing with patterns from prior workCHI
- 2025Authoring LLM-based assistance for real-world tasksIUI
- 2025Agent Base, a multi-agent orchestrator for coding agentsopen source
- 2024Mercura, AI quote and order automationYC W25
- 2024Dynamic abstractions workshopUIST
- 2023WorldSmith: multimodal world buildingUIST
- 2023Choice over control: writing with LLMsCHI
- 2022How to prompt? Prompting as an interaction problemarXiv
- 2022Beyond text generation: margin summaries for writersUIST
- 2022Suggestion lists vs. continuous generation on mobileMuC
- 2022SummaryLens: everyday text summarizationIUI
- 2022GANSlider: controlling generative models with slidersCHI
- 2021GestureMap: visual analytics for motion elicitationCHI
Work
Beyond Text Generation
An editor that keeps a short summary of each paragraph in the margin. Writers used the summaries to notice what their text actually said, then revised it.
SummaryLens
A small app: point your camera at a text, get a summary. A probe of what everyday summarization could feel like.
How to Prompt?
An early look at prompting as an interaction problem: what writing instructions for a generative model asks of people, and where it breaks down in creative work.
GestureMap
A tool that maps hours of full-body motion recordings into a 2D overview you can explore, compare and count.
Agent Base
A shared view of what a team's AI coding agents are doing: live sessions, short summaries, and enough context to jump into a teammate's conversation.
Choice Over Control
A study with 129 people writing with LLMs. Most preferred choosing between suggestions over instructing the model. Introduced the distinction between diegetic and non-diegetic prompting.
GANSlider
How people explore generative image models with sliders, and what small preview thumbnails change about that exploration.
Authoring LLM-Based Assistance
How people set up context-aware LLM assistance for tasks in their own surroundings. Built during my time at Meta Reality Labs.
Now
I'm building something new. It isn't public yet.
