Director, Model Behavior & Evaluation Systems
Blue Yonder · Paris
About Blue Yonder
Blue Yonder is the AI company for supply chain. Our platform helps the world's leading companies plan, fulfill, deliver, and operate more resilient supply chains across complex global networks.
We are building toward the autonomous supply chain: intelligent systems that can understand operational context, reason through tradeoffs, use tools, collaborate with people, and take action across real supply chain workflows.
About Autonomy Labs
Autonomy Labs' mission is to find the fastest possible path to an autonomous supply chain.
We build LLM agents, learning systems, model training pipelines, evaluations, simulations, and decision-making systems for some of the hardest problems in global supply chain. The work spans LLMs, agentic workflows, tool use, software automation, evaluation, post-training, optimization, and production engineering.
In short, we are having a lot of fun.
Your Mission
We are looking for a deeply technical Director of Model Behavior & Evaluation Systems to own the behavioral quality system for Blue Yonder's LLM agents.
Our agents are not generic chatbots. They are being trained to operate supply chain software: querying state, calling APIs, interpreting operational context, proposing actions, handling exceptions, asking for missing information, and helping users make decisions in complex enterprise environments.
Your mission is to make model behavior a product-quality system, not a collection of dashboards. You will define what "good" means for agents operating supply chain workflows, establish the release gates that determine when behavior is ready to ship, and build the feedback loops that turn traces, customer feedback, SME review, telemetry, red-teaming, and eval failures into model improvements.
This is a director-level technical leadership role. You will lead through systems, standards, people, and decisions. You should be close enough to model traces, evals, post-training, tool use, and customer workflows to make strong technical calls, while operating at the level of ownership boundaries, launch authority, roadmap sequencing, and team building.
The stack is real and close to the work. You should expect to operate around Python, PyTorch, Hugging Face Transformers and Datasets, NVIDIA NeMo RL, OpenAI Agents SDK, Langfuse, LLM evaluation harnesses, tool-calling traces, model checkpoints, reward and preference data, synthetic scenarios, experiment reports, and production observability. You do not need to be the person implementing every pipeline, but you do need the technical depth to challenge designs, read artifacts, understand failure modes, and guide senior engineers toward better systems.
What You'll Do:
Own the behavioral quality bar for Blue Yonder's LLM agents across customer-facing supply chain workflows.
Build and lead the model behavior and evaluation systems function across behavior specs, eval governance, SME review, release gates, regression coverage, and launch readiness.
Define launch criteria across operational correctness, tool-use accuracy, workflow completion, escalation quality, safe fallback behavior, refusal quality, consistency, and customer trust.
Establish evaluation authority so evals become release decision infrastructure, not just model-quality reporting.
Set the technical direction for behavior and eval infrastructure across Python eval harnesses, OpenAI Agents SDK workflows, Langfuse traces, LLM-as-judge workflows, deterministic checks, trace analysis, reward/report versioning, and model-candidate comparison.
Convert model traces, tool-call failures, SME feedback, red-team findings, telemetry, and customer-facing failures into behavior specs, eval requirements, training data needs, and model improvement priorities.
Partner with the reinforcement learning and post-training organization to turn behavior gaps into SFT data, preference data, reward criteria, curriculum, NeMo RL experiments, model-candidate decisions, and regression tests.
Partner with workflow, data, product, and domain experts to turn supply chain workflow truth into durable scenario coverage, rubrics, synthetic scenarios, eval datasets, and training data requirements.
Partner with agent architecture and product engineering teams to ensure prompts, tools, APIs, skills, system instructions, and product workflows express the intended model behavior consistently.
Review model traces, eval outputs, experiment summaries, dataset slices, reward reports, and post-training results closely enough to make informed launch and roadmap decisions.
Own customer and user behavior discovery for agent workflows: what users expect agents to do, explain, ask, verify, escalate, refuse, and act on.
Lead red-teaming and behavioral risk programs for hallucinated operational claims, incorrect tool use, overconfidence, poor escalation, prompt injection, unsafe recommendations, sycophancy, and unhelpful refusals.
Define the operating cadence for behavioral quality reviews, model release readiness, model cards, issue triage, regression management, and post-launch behavior monitoring.
Build a high-performing team or function around model behavior, evaluation governance, SME review systems, and behavioral launch quality.
Communicate model behavior strategy, quality tradeoffs, launch readiness, and residual risks clearly to executives, customers, product teams, engineering teams, and domain experts.
What We're Looking For - We want to talk if you:
Have led technical work on LLM products, AI agents, model behavior, model evaluation, post-training, AI alignment, or AI product quality.
Have built or led evaluation systems for open-ended LLM behavior, including eval datasets, graders, rubrics, scorecards, SME review loops, regression tests, or launch gates.
Have hands-on depth with LLM tool calling, function calling, API agents, workflow agents, software-operating agents, or enterprise automation.
Are technically fluent in the modern LLM stack, including Python, PyTorch, Hugging Face Transformers, Hugging Face Datasets, OpenAI Agents SDK, Langfuse, model checkpoints, tokenization, inference behavior, and experiment analysis.
Understand post-training workflows well enough to partner deeply with ML teams, including supervised fine-tuning, preference data, RLHF/RLAIF, reward modeling, NeMo RL, system prompting, synthetic data, and model launch evaluation.
Can reason about practical LLM behavior failures: hallucination, sycophancy, overconfidence, verbosity, refusal behavior, instruction hierarchy, prompt injection, tool-use failures, eval drift, and behavioral regressions.
Have enough technical fluency to inspect model traces, tool-call transcripts, eval failures, telemetry, datasets, experiment configs, and model-quality reports, then guide teams toward concrete fixes.
Have strong product judgment and experience translating user or customer needs into shipped product behavior.
Can lead cross-functional work across product, research, engineering, design, security, legal, customer success, domain experts, and executives.
Are comfortable making launch-readiness decisions under ambiguity, with explicit tradeoffs and evidence.
Are an exceptional written communicator who can turn ambiguous behavioral principles into clear guidelines, examples, rubrics, specs, and decisions.
Have experience defining model specifications, behavior guidelines, system instruction hierarchies, annotation rubrics, policy taxonomies, or model behavior standards.
Have managed, built, or strongly influenced senior technical teams.
Experience That Stands Out - Strong candidates may also bring experience in one or more of the following areas:
Owning model behavior, assistant quality, agent behavior, LLM product quality, or AI alignment work at a frontier lab, AI product company, or enterprise AI platform.
Scaling an evaluation, model-quality, or launch-readiness function from early experimentation into repeatable operating practice.
Working with NVIDIA Nemotron, vLLM, Ray, distributed evaluation, large-scale inference, or model-serving observability.
Applying AI agents to supply chain planning, warehouse management, transportation, logistics, operations research, retail, manufacturing, or enterprise workflow software.
Working with strategic enterprise customers on ambiguous, high-impact AI product behavior, safety, or launch-readiness decisions.
Publishing, open-sourcing, or internally leading notable work around model behavior, evals, agents, post-training, or AI product quality.
Technical Environment - You should expect to work around:
Python, PyTorch, Hugging Face Transformers, Hugging Face Datasets, and modern LLM post-training workflows.
NVIDIA NeMo RL, preference data, reward criteria, synthetic scenarios, model candidates, and experiment reports.
Tool-calling agents, OpenAI Agents SDK, API traces, workflow state, prompt and system-instruction behavior, and software-operating agent failure modes.
LLM evaluation harnesses, deterministic checks, LLM judges, SME trace review, grader rubrics, regression suites, and release gates.
Langfuse and production observability for model behavior: traces, telemetry, customer feedback, incident review, post-launch regressions, and model-quality reporting.
What Makes This Role Different? Model behavior is product behavior.
For supply chain agents, the difference between a useful system and an unsafe or frustrating one often comes down to judgment: when to ask for clarification, when to call a tool, when to explain a tradeoff, when to surface uncertainty, when to escalate to a human, and when not to act.
This role owns that judgment as an operating system: the standards, evidence, release gates, teams, and feedback loops that determine whether our agents are ready to operate in real customer workflows.
Equal Opportunity
All qualified applicants will receive consideration for employment without regard to race, color, religion, sex, sexual orientation, gender identity, national origin, disability, protected veteran status, or any other legally protected characteristic.
Our Values
If you want to know the heart of a company, take a look at their values. Ours unite us. They are what drive our success – and the success of our customers. Does your heart beat like ours? Find out here: Core Values
All qualified applicants will receive consideration for employment without regard to race, color, religion, sex, sexual orientation, gender identity, national origin, disability or protected veteran status.