Senior Director of Engineering, Generative AI

Roblox•San Mateo, CA

6d•$525,510 - $573,690•Onsite

About The Position

Roblox Build is our generative creation product, the platform where creators design, build, and publish 3D experiences. We are looking for a Senior Director of Generative AI to lead the Applied AI organization inside Build, responsible for turning state-of-the-art foundation models into high-quality, reliable creation systems at Roblox scale. This leader will own the full applied AI stack: model strategy and routing, model adaptation and fine-tuning, code generation (CodeGen), 3D layout generation (LayoutGen), and the evaluation science and infrastructure that tells us what actually works.

Requirements

10+ years of experience in machine learning and AI, with 5+ years in senior technical leadership roles overseeing applied AI or ML engineering organizations that have shipped generative AI systems into production at scale.
Deep technical grounding across multiple areas of applied AI, including LLM post-training (RLHF, DPO, distillation), model routing and adaptation, agentic systems, code generation, or evaluation science, with enough depth to recruit, challenge, and lead exceptional scientists and engineers in each area.
Demonstrated experience building and leading high-performing organizations that span both AI research scientists and production engineers, including hiring, developing talent, and making difficult personnel decisions.
Strong evaluation fluency, with a track record of building eval frameworks that predict production outcomes, designing benchmarks that measure what matters, and building the infrastructure to run experiments continuously at model, prompt, and routing level.
Experience operating cross-functionally in complex technical organizations, with the ability to influence and align across research, product engineering, and platform teams without relying on hierarchy.
Genuine interest in the creation problem. You find the specific challenge of AI-assisted 3D world creation and domain-specific code generation compelling, not just as a vehicle for building AI systems.

Nice To Haves

A builder and a shipper who cares about getting things into production and measuring whether they actually work, not just publishing promising results.
Technically credible across the stack. You can engage deeply with scientists on evaluation methodology, data quality, and model behavior, and equally deeply with engineers on architecture, inference latency, and reliability.
Decisive under ambiguity. You make principled tradeoffs between rigor and velocity, know when good enough is good enough, and can articulate your reasoning clearly.
Direct and honest. You give clear feedback, surface problems early, and create an environment where your team can do the same.
Mission-driven. You care about the people building on Roblox, and you connect the technical work of the Applied AI org to the experience of the millions of creators using the tools.

Responsibilities

Own model strategy and routing for Build. Design and build an intelligent model layer that selects the right model for each creation task based on quality, capability, latency, cost, and safety, leveraging both frontier models and Roblox-adapted open-source models.
Lead model adaptation across the Applied AI org, including fine-tuning, distillation, synthetic data generation, human feedback pipelines, and preference optimization for Roblox-specific creation tasks such as Luau code generation and 3D scene understanding.
Drive CodeGen capabilities forward by building AI systems that understand creator intent, reason across multi-file Roblox experiences, execute tools, and reliably make complex changes to existing games, going well beyond code completion.
Build LayoutGen intelligence by developing the AI capability that turns a creator's intent into coherent 3D scenes, including object selection, spatial reasoning, placement, aesthetic quality, and iterative editing.
Own evaluation as a core technical discipline. Build the benchmarks, quality metrics, judge methodologies, experiment infrastructure, and human evaluation pipelines that answer whether model, prompt, or routing changes actually improve creator outcomes.
Close the feedback loop between model development, evaluation, and production, ensuring that quality signal flows continuously from creator outcomes back into training data, model updates, and release decisions, rather than operating as separate silos.