The Fog
A word is never alone.
Enter a word, discover its neighbours, and select another word to continue. Slowly shifting fog creates a sense of discovery; a clear typographic hierarchy and a pause control keep the map readable.
The Fog & The Gravity
Making language
feel like a landscape.
How can we make the way AI represents language tangible and explorable?
I developed two interactive prototypes around word embeddings—numerical representations learned from patterns of word use. My focus was connecting model behaviour with interface decisions: translating vector similarity into spatial relationships people can navigate, question, and compare.
The Fog uses pretrained GloVe embeddings to let people follow associations from one word to the next. The Gravity uses separate Word2Vec models trained on Pride and Prejudice and Jane Eyre to explore how different texts shape those associations. Across both, I worked between model outputs and visual feedback, deciding what the interface could meaningfully reveal and where it needed to acknowledge approximation.
- Role
- Creative Technologist
- Skills
- AI prototyping, data visualisation & interaction design
- Time
- 2026
- Tools
- Python · Gensim · JavaScript · Canvas
The Works
One visual language, two ways to explore: follow an association, then compare how context changes it.
The Gravity
A word belongs to its world.
Compare a shared vocabulary in Pride and Prejudice and Jane Eyre. Linked hover connects a word across both views, while colour and type size reveal different associations and frequencies.
Design Research
The Fog
Pretrained GloVe embeddings turn a word into a neighbourhood of associations. I used proximity to suggest related usage and made each word a path to explore. Colour creates atmosphere rather than encoding similarity.
The Gravity
Two Word2Vec models learn separately from two novels. I placed the same words in linked views, making it possible to compare their relationships while keeping each model’s context distinct.
Same words, different contexts
Both maps share a vocabulary: 18 study words for a focused comparison, or neighbours drawn from both models for discovery. Keeping the words consistent makes differences in their relationships easier to follow.
When colour becomes information
When colour became purely decorative, I asked for its semantic meaning back. Each word now follows its closest anchor—love, reason, or money—within each novel’s model.
Affection shifts from love in Austen to reason in Brontë. Type size shows frequency adjusted for book length, so a longer text does not automatically make words larger.
- Love
- Reason
- Money
These are exploratory maps: nearby words are not necessarily synonyms, 2D distances are approximate, and scores from separately trained models are not directly comparable.
Interaction Development
Offer a way in
The early Fog opened with an ocean map. I replaced that predetermined result with an empty cloud, then added click-to-discover: visitors can enter a word or let a random query start their exploration.
One word becomes the next question
Selecting a neighbour queries the same saved model again; it does not retrain it. Here, city leads to metropolitan, then downtown. I made each result a new entry point, so exploration can continue without repeatedly typing.
Two contexts, one entry point
The intertwined clouds looked like one object, obscuring the idea of two distinct texts. I separated them around a shared centre and kept one visible search field: a single query opens both models for comparison.
The attraction is a designed metaphor for two texts in dialogue, not a force calculated from the novels.
Keep the evidence steady
Because position carries meaning, the words settle into fixed locations after the opening transition. I kept movement in the fog and linked hover across both maps, letting readers track the same word and inspect its associations.
Working between model and experience
With an AI coding agent supporting implementation, I directed how model outputs became an experience—checking the meaning of distance and colour, reviewing working prototypes, and refining how people explore the results.