Full-Time
AI Engineer
Full-stack
Taste Labs
11-50 employees
Design taste data for training AI models
Compensation Overview
$175k - $275k/yr
San Francisco, CA, USA
In Person
 AI & Machine Learning (1) |
|---|
Get referred to Taste Labs
See people who can refer or advise you
- Experience building at early-stage companies from pre-seed to Series C and operating effectively in ambiguity.
- Experience shipping agent systems and building with large language models through professional work, open source, or serious personal projects.
- Genuine curiosity about taste and comfort working on hard, nuanced, and undefined problems.
- Creative problem-solving and the ability to invent infrastructure for a new problem space.
- Build AI systems for taste evaluations, tooling, data collection, application programming interfaces, and reinforcement learning environments, including agent architectures, data pipelines, and product surfaces.
- Craft agent harnesses, memory systems, and self-improvement loops.
- Design evaluation pipelines and synthetic data generation systems.
- Design evaluations and grading systems for unverifiable domains.
- Create embedding and retrieval infrastructure that scales to millions of requests.
- Build crawling and scraping systems for visual data across the web.
- Set up inference serving and application programming interfaces for client-facing products.
- Develop tooling and infrastructure that makes systems reliable and fast.
- Ship internal tools for data operations and external tools for expert annotators.
- Build gamified product experiences including taste quizzes, leaderboards, and reward flows.
- Open source contributions or personal projects that demonstrate curiosity and building.
- Background at creative companies such as Figma, Notion, Canva, Adobe, Runway, or at companies with strong indexing and crawling or data capabilities.
Taste Labs builds the data and infrastructure layer that AI systems use to judge subjective quality, starting with design. The company measures, classifies, and codifies aesthetic judgment into data models that systems can train on, then supplies the context, evaluation, and verification tooling that sits on top. It works with frontier labs on post-training to improve taste and design capability, and with app developers, coding agents, and creative-tool companies checking the design work their models produce. Unlike benchmarks that score correctness, it targets the harder question of whether output is good, on-brand, or right for a particular person. The company's goal is to give models reliable judgment by turning taste into training data instead of leaving it to prompt engineering.
Company Size
11-50
Company Stage
Seed
Total Funding
$18.5M
Headquarters
San Francisco, California
Founded
2025
Get referred to Taste Labs
See people who can refer or advise you
Simplify's Take
What believers are saying
- June 2026 launch sparked strong attention around ending AI slop and subjective quality.
- Frontier labs and app companies now pay for better design, brand alignment, and evaluation.
- Amplify’s June 16, 2026 investment validates demand for post-training taste infrastructure.
What critics are saying
- Taste is subjective, and customers can replace its rubric with cheaper prompts.
- OpenAI, Anthropic, and design tools can build in-house preference data quickly.
- If AI-generated style becomes commoditized, Taste Labs loses the scarcity of human taste.
What makes Taste Labs unique
- Taste Labs sells judgment data, rubrics, and evaluation environments for subjective AI quality.
- Thais Castello Branco launched June 16, 2026 with Amplify and CRV backing.
- It starts in design, using 1,000 vetted tastemakers instead of crowdsourcers.
Help us improve and share your feedback! Did you find this helpful?
Benefits
Remote Work Options
Flexible Work Hours
Company News
AI weekly: the war on slop. The AI conversation flipped from volume to quality, and every fix is human judgment The war on slop. "AI agents" came up in 33% of last week's business podcast episodes, up from 26% a month ago, the biggest mover of any AI term Meet Bri track. The agents are everywhere. And the conversation about them just flipped from volume to quality. The word for the failure mode is slop, and last week marketing, venture, engineering, and research all declared war on it. Product leader Hillary Gridley called internal AI slop an epidemic and described the failure spiral: people stop questioning AI output, "the people get worse, that makes the systems worse, and then you get into this sort of slop doom loop." Her diagnosis is that slop is a leadership failure, not a tool failure. If nobody defines what good looks like, "You can't be surprised when that quality starts slipping." Venture is now funding the fix. Thais Castello Branco just raised an $18.5 million seed for Taste Labs, a company that pays roughly a thousand designers and critics to teach models what good looks like, because models trained to converge on the most likely answer are structurally biased toward average. Jason Calacanis put the consumer version bluntly: "The people who build these things have no taste. Let's be honest. They don't have taste." The platforms are moving too. LinkedIn shipped a report-slop button. Substack shipped anti-AI detection tools. Engineering had the same argument with different words. The "software factory" push, teams cutting code review to push agentic code straight to production, runs into a benchmark gap: a benchmark "will tell you whether or not the code works, but it doesn't actually tell you anything about whether or not it's maintainable." The research world gave the phenomenon its scientific name: cognitive debt. Researcher Margaret-Anne Storey watched student teams build MVPs in weeks with AI, then hit a wall when nobody understood what they had built. "With cognitive surrender, you're surrendering not just your understanding of it, but also your ability to learn." Her prescription is strategic friction, borrowed from a colleague's line that "cars have brakes" so they can go faster. Here is the convergence worth noticing. Nobody in these four conversations is mad at the models. They are mad at the humans who stopped supervising them. Every prescription on offer, rubrics, paid tastemakers, code review as back pressure, deliberate friction, is a mechanism for putting codified human judgment back in the loop. First drafts became free last year. The value moved to whoever sets the standard. Sources: Hillary Gridley (ex-WHOOP), Marketing Against the Grain, "If Your Team Is Producing AI Slop, Here's How To Fix it," Jul 28, 2026. Thais Castello Branco (Taste Labs), This Week in Startups, "Why AI has no taste and how to fix it," Jul 31, 2026. Ben Lloyd Pearson and Andrew Zigler (LinearB), Dev Interrupted, "The rise of software factories, the fall of first drafts, and the hidden tax holding back your agents," Jul 31, 2026. Margaret-Anne Storey (University of Victoria), Tech Lead Journal, "From Technical Debt to Triple Debt: The Hidden Cost of AI-Generated Code," Jul 27, 2026. The slop backlash is a judgment crisis: rubrics, paid tastemakers, code review, and deliberate friction all put codified human standards back in the loop.
By Veronica Orellana, Reid Christian and Mia Krishnamurthy