AI Engineer, RL

Tastelabs; San Francisco, Northern (CA, KY); Hybrid; USD 120,000 - 190,000; Full time

Benefits offered by this job

Open source contributions

Job summary

Taste Labs in San Francisco is building the data and infrastructure layer for taste, advancing evaluation methods for subjective design tasks and enabling frontier models to perform with context and verifiability.

You will design tasks that capture elements of taste, build RL infrastructure, and work with research teams to craft environments that improve model behavior. A startup mindset and backend skills are essential.

Qualifications

• Experience designing evals and grading schemes for non-verifiable outputs.
• Experience building RL environments and backend systems for data pipelines.
• Strong backend development skills and ability to ship in a startup setting.

Responsibilities

• Research grading methods and rubrics for subjective domains.
• Design tasks capturing elements of “taste” and design capabilities.
• Build agent harnesses and context layers for evaluation pipelines.
• Work on scalable RL infrastructure and data pipelines.
• Collaborate with internal research teams on training pipelines and evaluation methods.

Skills

Eval design

RL environments

Backend development

Startup experience

Tools

Python

Job description

About the company

Taste Labs is building the data and infrastructure layer for taste. Our goal is to end AI slop. To make AI feel right, not just be correct. We raised $18.5M in seed co-led by Amplify and CRV, and most frontier labs are already customers. AI has nailed objective domains and can generate anything. The hard part left is judgement: what fits, what feels like you, what's actually GREAT. We're turning that into something measurable, starting with design. We do it on two sides: building the post-training data and RL environments that teach taste to frontier models, and the context and verification tools agents need to produce work that's more creative, more on-brand, more right.

If that problem excites you, you'll like it here!

About the role

How do you judge something non-verifiable, like design? To solve subjective domains, we have to solve how to grade the difference between slop and great without the need for humans in the loop.

Types of problems you’d work on

• Research different grading methods and rubrics
• Design unique tasks that can capture elements of “taste” and design capabilities
• Build agent harnesses and context layers
• Work on scalable RL infra
• Work with internal research teams on our training pipelines
• Work with the top frontier labs on how to craft environments to improve frontier models

What matters to us

• Experience in building evals, RL environments, ML or post-training. Strong backend experience as well.
• You like ambiguous, hard, creative problems, and want to make subjective domains verifiable
• You’re a team player and have startup DNA: you move fast, adapt, no such thing is ‘not in scope’, you like to take ownership of things. You’ll be given a lot of room to run.

Bonus points

• Open source contributions or personal projects that show you build things because you're curious
• Background at creative companies (Figma, Notion, Canva, Adobe, Runway, etc.) or companies with strong index building/crawling (e.g. Firecrawl, Brave, Luma, Pika) or data (Mercor, Surge, etc.)