Datasets lack representation; Captions lack semantic context; Generative models don’t understand user intent
We are building multimodal encoders that bridge these gaps
Captions are what give meaning to pixels, yet most of visual culture has been digitalised or produced with poor captioning practices, resulting in text descriptions that can’t provide the meaning, context and purpose required for training representative and pluralistic AI.
Digital assets, especially scraped and user-generated media, lack standardized semantic structure and robust metadata; most digital media was tagged for commercial web indexation, not for conveying complex composition and meaning structures
To make these assets useful for model training, developers must artificially enrich it. This process creates a scenario where foundational datasets become open to subjective interpretation and systemic appropriation.
To address metadata debt, captions are outsourced to gig workers, and increasingly, to AI models, neither of which have the contextual knowledge to describe agency, relationships, causality and uncertainty across millions of different assets.
Consequently, foundational training data becomes generic, biased and stripped of the context that gives it meaning.
Based on captions, encoders then organise images into embedding spaces that determine which similarities and differences models can recognise.
When captions fail to establish how similar images express different beliefs, intentions and relationships, encoders compress reality into patterns, relegating less represented interpretations to outliers and making culturally significant distinctions harder for downstream models to recognise.
Encoders are the medium through which captions become
context about the world.
We are curating a contextually captioned dataset and building a multimodal encoder trained to preserve alternative, contradicting and divergent structures of meaning. We focus on three critical challenges:
We are developing a computational protocol that encodes how different knowledge systems organise the same visual evidence.
This allows us to treat context, agency and causality as explicit dimensions of training data, turning interpretations that would normally be classified as outliers into distinct learning signals.
Using these datasets, we are developing multimodal encoders that preserve differences in salience, association and causality within embedding spaces.
These encoders carry the learned distinctions into new images and descriptions, turning semantic plurality into a learned capability that shapes how AI interprets unfamiliar content.
Our goal is to make these datasets, encoders and benchmarks available for researchers and developers to train, adapt and evaluate their own systems.
This gives developers the infrastructure to train and evaluate AI on competing interpretations of the same world, making disagreement a source of model capability instead of an inconsistency to eliminate.