At Pareto.AI, we’re on a mission to enable top talent around the world to participate in the development of cutting-edge AI models.
In coming years, AI models will transform how we work and create thousands of new AI training jobs for skilled talent around the world. We’ve joined forces with top AI and crowd researchers at Anthropic, Character.AI, Imbue, Stanford, and University of Pennsylvania to build a fair and ethical platform for AI developers to collaborate with domain experts to train bespoke AI models.
Pareto builds human training data pipelines for frontier AI labs. As a Data Delivery Lead, you sit at the center of that work — owning the architecture, execution, and continuous improvement of complex data collection and evaluation workflows from first scoping call to final delivery.
This is a technical operations role, not a project management role. You'll be expected to read code, deep dive data, reason about LLM internals, design evaluation frameworks, and — increasingly — deploy and iterate on AI agents to automate the work your pipelines do today. We're actively building toward a model where agentic systems handle quality gates, expert routing, and output review, and DDLs are the people designing and operating those systems.
You'll work directly with AI researchers and technical program managers at our client organizations, own delivery against model performance benchmarks, and lead a team of project managers who handle day-to-day execution tracking.
What you'll do
Pipeline architecture Design end-to-end data collection and evaluation pipelines for RLVR, RLHF, SFT, red-teaming, and model evaluation workflows. This includes expert sampling strategy, annotation schema, rubric structure, inter-rater calibration, and QA system design. You'll be expected to prototype novel workflows quickly, identify architectural risks before launch, and make tradeoff decisions with confidence. You’ll also need to understand how agents interact with tools to solve expert-driven tasks, and you’ll need to communicate with engineering to ensure the environment is built accordingly to enable such tasks.
Agentic system deployment Build, test, and iterate on AI agents that automate pipeline tasks — quality gate review, expert matching, output flagging, throughput anomaly detection. You'll work closely with our engineering team to scope agent capabilities, write the prompts and evaluation logic that make them reliable, and monitor their performance in production. This is a growing part of the role; comfort with agentic tooling (LangChain, DSPy, custom tool-use frameworks, or equivalent) is a meaningful differentiator.
Quality systems Define data quality standards across annotation, evaluation, and expert output review. Design and run audits using inter-rater reliability metrics, calibration sets, and statistical sampling. You'll be responsible not just for catching quality issues but for building systems that prevent them — automated checks, structured output validation, and model-assisted review layers where appropriate. Aside from programmatic quality testing, you’ll be responsible for spot-checking tasks & understanding what makes a datapoint meaningful & high quality. The ability to do so, and translate your findings into clear feedback and expert guidelines is particularly helpful for this role.
Client interface Engage directly with AI researchers, TPMs, and PMs at our client organizations. Translate research-driven requirements — evaluation rubrics, domain coverage targets, latency constraints, benchmark specifications — into operational workflows. Communicate pipeline performance clearly, escalate technical risks early, and contribute to project scoping and pricing decisions.
Research integration Stay current with developments in LLM post-training, evaluation methodology, and data tooling. Evaluate new approaches — model-assisted annotation, structured output formats, automated calibration methods — and integrate them into active pipelines where they improve quality or efficiency. Understand what method applies to what domain and project, and work towards implementing it accordingly.
What you'll need
Proficiency in Python and SQL for data manipulation, pipeline monitoring, and quality analysis — you should be comfortable writing light scripts to parse formats, run statistical checks, and build lightweight tooling
Working knowledge of LLM internals: RLHF/SFT training loops, how prompt structure affects output distribution, RL environment setup qualities (tool use) for agentic data collection / eval projects.
Hands-on experience with at least one agentic or LLM workflow framework (LangChain, DSPy, AutoGen, direct tool-use via API, or equivalent)
Demonstrated ownership of a data or ML pipeline from scoping through delivery — including quality design, not just throughput tracking
Strong written communication: you'll write technical guidelines and rubrics that distributed expert workers follow accurately, and you'll brief senior researchers on pipeline performance
Comfort operating with ambiguity in a fast-moving environment where model requirements shift and client priorities evolve
You'll stand out if you have
Direct experience with RL environment data pipelines, evaluation framework design, and red-teaming workflows
Background in data engineering, ML research support or equivalent
Experience designing or operating agentic systems in a production or near-production context
Familiarity with inter-rater reliability methods, calibration set design, and annotation quality frameworks
Prior client-facing or technical program management experience in an AI/ML-adjacent context
Prior experience on scoping or driving projects with fuzzy upfront specs or evolving requirement. This is a high-ownership position where you are expected to take charge and lead, not simply follow project guidelines.
What we value in candidates
We care less about credentials than about demonstrated ability to own complex technical work and build toward better systems. A strong background in software engineering, data science, or ML research is the most common path into this role, but we've also seen excellent DDLs come from ML operations, computational linguistics, and applied research support. What matters is that you can read a dataset, reason about what's wrong with it, write code to fix it, and design a workflow that prevents the problem next time.