Taste Labs is building a data and infrastructure layer aimed at the parts of AI output that are hard to score with a unit test, things like whether a generated design feels on-brand, original, or visually right. The company is launching out of stealth with an $18.5M seed round co-led by CRV and Amplify Partners, targeting the flood of generic outputs often called slop by supplying preference datasets, rubrics, and evaluation environments to frontier labs and design-focused applications rather than broad consumer tools . The intended buyers are frontier model teams that need post-training data and RL environments for subjective tasks, plus generative AI and agent companies that need context and verification tools so their outputs land closer to what a human would actually ship.
The team is starting with design as the first vertical, using it as a concrete test bed for codifying and searching over what "great" means when there is no single correct answer. That focus matters now because the wave of agentic and generative products hitting market is exposing how thin current evals are for taste-sensitive work, and incumbents in MLOps and evaluation are largely built around objective benchmarks rather than aesthetic judgment.
Taste Labs is led by founder and CEO Thais Castello Branco, who was previously on the founding team at Exa, and the company is based in San Francisco and hiring across research, engineering, and product.
Comments (8)
Sign in to join the discussion.
N
Nadia KrollJun 16, 2026
Ending AI slop is a bold promise from a team that just dropped a launch video with a slo-mo macaron shot. Respect the commitment to the bit.
O
Olu A.Jun 16, 2026
Taste is downstream of conviction, and conviction does not fit in a JSON schema. Curious how you encode that.
M
Mireia VidalJun 16, 2026
Was already pulling the term sheet out of my drafts when I saw the co-leads. Anyone needs a warm intro to a designer who actually uses the product, my DMs are open.
D
dev k.Jun 16, 2026
Cool, but where are the docs? Give me a /score endpoint, a rate limit page, and a webhook for when my output gets flagged as slop.
T
Tomasz BielskiJun 16, 2026
The market for 'taste' shrinks the moment models get good enough that mid users stop noticing. You are selling discernment to people who picked Comic Sans on purpose.
H
henrikJun 16, 2026
Building in the eval space too and genuinely happy to see this framing land. Slight rivalry energy but also, finally someone willing to say the quiet part.
Y
Yuki M.Jun 16, 2026
The wordmark has a rogue space between 'the' and 'Taste' in your tweet copy. If you are the taste layer, that one is going to haunt the replies.
R
Rashida OkaforJun 16, 2026
Before I forward this to our platform team: SOC 2, data residency in EU, and can you sign a DPA before Q1? Procurement does not care about ending slop.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Taste Labs
Decoding subjective domains the Taste infra layer for AI
Decoding subjective domains the Taste infra layer for AI
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Taste Labs is building a data and infrastructure layer aimed at the parts of AI output that are hard to score with a unit test, things like whether a generated design feels on-brand, original, or visually right. The company is launching out of stealth with an $18.5M seed round co-led by CRV and Amplify Partners, targeting the flood of generic outputs often called slop by supplying preference datasets, rubrics, and evaluation environments to frontier labs and design-focused applications rather than broad consumer tools . The intended buyers are frontier model teams that need post-training data and RL environments for subjective tasks, plus generative AI and agent companies that need context and verification tools so their outputs land closer to what a human would actually ship. The team is starting with design as the first vertical, using it as a concrete test bed for codifying and searching over what "great" means when there is no single correct answer. That focus matters now because the wave of agentic and generative products hitting market is exposing how thin current evals are for taste-sensitive work, and incumbents in MLOps and evaluation are largely built around objective benchmarks rather than aesthetic judgment. Taste Labs is led by founder and CEO Thais Castello Branco, who was previously on the founding team at Exa, and the company is based in San Francisco and hiring across research, engineering, and product.
Comments (8)
Ending AI slop is a bold promise from a team that just dropped a launch video with a slo-mo macaron shot. Respect the commitment to the bit.
Taste is downstream of conviction, and conviction does not fit in a JSON schema. Curious how you encode that.
Was already pulling the term sheet out of my drafts when I saw the co-leads. Anyone needs a warm intro to a designer who actually uses the product, my DMs are open.
Cool, but where are the docs? Give me a /score endpoint, a rate limit page, and a webhook for when my output gets flagged as slop.
The market for 'taste' shrinks the moment models get good enough that mid users stop noticing. You are selling discernment to people who picked Comic Sans on purpose.
Building in the eval space too and genuinely happy to see this framing land. Slight rivalry energy but also, finally someone willing to say the quiet part.
The wordmark has a rogue space between 'the' and 'Taste' in your tweet copy. If you are the taste layer, that one is going to haunt the replies.
Before I forward this to our platform team: SOC 2, data residency in EU, and can you sign a DPA before Q1? Procurement does not care about ending slop.