Thais Castello Branco
Conference affiliation: Founder · Taste Labs · 2026
Related work and projects
- A manifesto for judgement ↗tastelabs.com · Source ↗
Thais Castello Branco is the founder and chief executive of Taste Labs, an AI research and infrastructure company building ways to evaluate and improve creative work, starting with design. Thais focuses on the gap between generating something that functions and making something that fits: a pet-shop website and a finance-company website should not converge on the same design simply because an agent can build both.
On this page
Her approach combines model training with tools that give agents context, inspiration, and feedback. Better design models are only part of the answer; the system also needs to understand what a particular user wants to make.
From consumer brands to AI products
Thais’s earlier work involved developing products and their identities. At Ambev, she worked in brand marketing. At Kraft Heinz, she helped launch brands through its Springboard startup incubator, worked on Claussen’s brand repositioning, and developed new products. In 2023, she led growth and marketing at Hubla, a platform for creators selling digital products, where her responsibilities included product positioning, branding, and pricing.
She also co-founded Creio, a Catholic community app, serving as chief product officer in 2023–2024 and subsequently as an investor. Her career history records these roles alongside participation in the Latitud Fellowship from January 2023 to August 2024. Latitud’s later investment account dates its first meeting with her to 2024.
Thais joined Exa’s founding team in August 2024. She managed and launched Websets, led data partnerships, and built the marketing function. Those responsibilities covered product management, commercial relationships, and bringing the product to an audience; they do not imply sole authorship of Exa’s technology. She left Exa in October 2025 and founded Taste Labs that month.
Latitud wrote Taste Labs’ first check. The company emerged from stealth in June 2026 with an $18.5 million funding round. Taste Labs works with frontier-model teams on evaluation, post-training data, and reinforcement-learning environments, while also developing tools for creative applications and agents.
Measuring quality without averaging away taste
Thais treats taste as an acquired skill: people build a repertoire, recognize patterns, develop a point of view, and learn what to reject. She argues that expecting every user to acquire that expertise in every creative domain is unrealistic. Systems should help non-experts articulate their intent and create better work, while preserving the distinctions experts have learned to make.
Three decisions shape that work:
- Turn brand identity into verifiable criteria. A request for great design leaves the target vague. An established brand supplies more concrete properties: colors, typography, spacing, motion, and texture. In Ending AI Slop, Thais uses Reducto’s website to illustrate how those properties could become criteria for a reinforcement-learning task. An agent would create a new page and be assessed against the brand’s components, allowing a different composition without rewarding mere imitation. This is an illustrative task design, rather than a reported training result. She cautions that model-based judges can hallucinate or reward shortcuts, making the choice of what to measure central to the task.
- Preserve meaningful expert disagreement. Disagreement about alignment may expose a defect; disagreement about minimalist versus maximalist aesthetics may reflect legitimate preferences. Thais favors retaining who prefers what and in which context, rather than pooling judgments into an average that satisfies neither audience. She also emphasizes specific criticism tied to a visual element or its corresponding code. Connecting a designer’s observation to the component it concerns gives a model clearer information than a preference vote on an entire page. She advocates carefully constructed data over annotation volume, while acknowledging that Taste Labs does not always receive a complete account of how its data affects a laboratory’s models.
- Make creative departures intentional. Familiar outputs can be a poor target for design and writing, but novelty still needs direction. In Training Taste, Thais describes a pitch deck that breaks selected conventions while retaining the expectations that make it useful as a pitch. Increasing sampling randomness alone cannot make that choice. The system needs context about the category, audience, and purpose to decide where variation helps.
These distinctions determine which method to use. Properties that can be checked programmatically lend themselves to verifiable tasks; questions of style, audience, and changing preferences require contextual human judgment. Her manifesto on judgment accompanies this work on subjective quality.
Taste Labs has also investigated how to detect repetition. In Training Taste, Thais describes the team analyzing more than two million websites from roughly a decade of web history and comparing historical designs with generated sites. The team extracted features such as palettes, typography, and layouts, then trained small classifiers to recognize recurring patterns. Combining those signals helped identify the repetition associated with generic AI output. Thais reports that the approach outperformed model-judge methods in that research, without giving numerical performance results in the recording. The finding concerns those design patterns and comparisons, rather than a universal measure of artistic merit.
Giving agents a brief and corrective feedback
Thais places considerable weight on inference time because that is when an application can clarify a user’s intent. A stronger model still needs to distinguish an existing business protecting its identity from someone creating a new business with only a loosely described style in mind.
Taste Labs’ Brand API addresses both situations. In her public product announcement, Thais describes three functions:
- Extract: turn a website URL into a structured brand system an agent can follow.
- Search: retrieve inspiration matching a requested direction, such as a dreamy brand, for users without an existing identity.
- Verify: critique generated work against the intended brand, identify departures, and suggest corrections.
The workflow reuses decisions designers have already made instead of asking an agent to infer an identity from a vague instruction. In Training Taste, Thais compares a slide deck generated for The General Intelligence Company of New York with one produced using Taste’s brand extraction. She presents the latter as a closer match to the original identity, including its details. The demonstration illustrates the intended use of the tool; it is not a controlled benchmark.
On August 18, 2026, Thais announced a Taste Labs research fellowship for investigations into model creativity and human preferences, offering compute, custom data, access to designers and researchers, and a stipend.
Read the topics behind these talks
2 conference talks
Loading more talks and highlights…

