AI Product Experiments

01 / The Friction & Problem

Most AI software is still being built as static legacy web forms with an awkward chatbot sidebar tacked on. This series of experimental prototypes asked: what does a graphical interface look like when the computational model is a co-author, not merely an asynchronous chatbot?

02 / The Core Concept

A suite of interaction primitives: streaming canvas editors, intent-driven direct manipulation, visible probabilistic confidence rings, and instantaneous reversible time-travel debugging.

03 / Research & Process

• Systematic literature review of modern Human-Computer Interaction (HCI) and LLM latency research
• Engineering 6 working browser prototypes exploring different token-streaming patterns
• Weekly critique sessions with senior interaction designers and ML research engineers
• Public write-ups and viral demo releases exploring failure states as creative material

04 / Key Design Decisions

• A canvas workspace where text doesn't dump into a chat bubble, but streams dynamically into contextual blocks.
• A transparent 'Why this response' inspector that surfaces citation provenance and token certainty.
• An atomic undo engine: every AI generation has a branched timeline that can be rewound or forked.
• Graceful degradation: visual cues that communicate model ambiguity before the user relies on a hallucinated premise.