Thais Castello Branco is founder and CEO of Taste Labs, which develops evaluation and feedback infrastructure for creative AI. Previously on Exa’s founding team, she focuses on helping models distinguish generic output from work suited to a particular user and context.[1][3]
Start with Training Taste →
Links
Taste Labs
X Founder and CEO · Taste Labs
Career
Latitud’s investment account describes meeting Castello Branco through its fellowship in 2024, her subsequent work at Exa and her departure to build Taste Labs. The company works with model developers and application teams on the judgment involved in design and other creative tasks. Her background in product and communication informs the distinction between generating an artifact and making one that fits its purpose.[3][2]
Contributions
In her Training Taste talk, Castello Branco distinguishes improving base-model creative capabilities from helping an application understand a particular request. The latter includes context, relevant inspiration and feedback on the produced design. Repeated evaluation can refine the artifact against its intended direction instead of judging only whether code runs.[5]
Ideas
Her manifesto describes taste as a repertoire built through exposure, observation, a point of view and selection. The engineering difficulty is that creative work lacks one universal correct answer: a strong result for one audience may be wrong for another. She argues that better models need both richer evaluation and context about the person, brand and task, rather than a single average aesthetic.[4]
Tools, writing & talks
Writing, research & talks
Start here
Training Tasteai.engineerA talk on creative evaluation, user context and feedback for design agents.
Start with the examples contrasting technically functional output with contextually suitable design. No model-training background is required; distinguish the proposed evaluation approach from a universally agreed definition of taste.
Speaker[5]Taste Labs mined features from two million sites to train small classifier probes that each detect one slop signature, then stacked them; the combination reportedly beats asking a model to judge aesthetic quality directly.
Catalog workOriginal source- A manifesto for judgementtastelabs.com
An argument for treating taste as a learnable, context-dependent skill.
Author[4]