Job Description

At Roblox, we're on a mission to empower all the creators of the world to build and share amazing immersive experiences. As a member of the Roblox Studio team, you’ll be working on the most popular tool for game development in the world on a platform with over 80M daily active users. Our Studio team is seeking an exceptional Director-level technical engineering lead to push the boundaries of AI-assisted game development.

The Core Problem:

Foundation models are powerful generalists, but unlocking their true potential requires deep specialization. At Roblox, we're building an AI Assistant to fundamentally change how millions of creators build immersive experiences. A critical component is generating high-quality, context-aware Lua code. This isn't about simple boilerplate; it's about understanding user intent, the nuances of the Roblox API, and the structure of complex game logic. Making a large language model genuinely useful – truly assistive – for programming in a specific, massive ecosystem like Roblox is a frontier challenge in applied AI.

As a Technical Director (IC role) for Roblox Assistant, you’ll own the overall technical vision for machine learning initiatives toward a future where the highest quality games can be created dramatically faster than is possible today and the strategy and execution for specializing our LLMs for code generation via post-training.

What We Need to Solve:

• Deep Specialization: How do we effectively adapt frontier LLMs to master the intricacies of Lua within the Roblox engine? 
• Data is King (and Hard): We have vast amounts of code, but transforming it into high-signal training data for SFT, preference tuning (RLHF, DPO, etc.), and robust evaluation is a complex machine learning and engineering challenge. We need to pioneer sophisticated quality signals and data mixing strategies.
• Beyond Instruction Following: Moving past simple prompt-response towards genuine reasoning and problem-solving for complex coding tasks. How do we improve the model's ability to handle ambiguity, plan multi-step and multi-file code generation, and debug?
• Rigorous Evaluation: Defining and implementing meaningful evaluations that capture true coding utility, beyond surface-level metrics.

You Have:

We need someone who lives and breathes improving LLM output quality, specifically for pushing performance in challenging, specialized domains.

• You are obsessed with the post-training loop: You have deep, hands-on experience and intuition for techniques like Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO), data curation/mixology, and building robust evaluation suites. You understand the subtle interplay between data quality, model scale, and tuning strategies.
• You have competed and won at the highest levels: You have a proven track record of taking state-of-the-art foundation models and significantly improving their capabilities on specific, difficult tasks – ideally code generation, but deep expertise demonstrated in any complex domain (e.g., mathematics, reasoning, native multimodal output, complex instruction following) is highly relevant. You've likely led teams of 25+ engineers pushing benchmark performance at leading AI labs or internal efforts within major tech companies.
• You are deeply technical and hands-on: While you will lead technical direction, you aren't afraid to dive into the weeds to diagnose tricky model behavior, analyze failure modes, or prototype new data processing techniques. You understand the mechanics, not just the concepts.
• You think rigorously about data: You appreciate that progress is often gated by data quality and effective curation. You have experience designing and scaling data annotation efforts and developing signals to identify high-quality data for specialized domains.
• Qualifications: A strong academic background (MS/PhD preferred) in CS, ML, or a related quantitative field.
  • Proven track record with 15+ years of experience including deep expertise in frontier LLM post-training .
  • Fluent in ML frameworks (like PyTorch) and data processing tools (like PySpark).

You Will:

• Define and drive the technical roadmap for improving Roblox Assistant's code generation and reasoning capabilities through advanced LLM post-training.
• Lead the design and execution of experiments in fine-tuning, preference alignment, and data strategy.
• Establish rigorous evaluation methodologies to measure true progress in coding assistance.
• Work closely with ML engineers, providing technical guidance and mentorship focused specifically on model improvement techniques.
• Stay at the absolute forefront of LLM research and apply relevant advancements to our specific challenges.
• Contribute to the core intellectual challenges of making AI a powerful, reliable tool for millions of software creators.

Why Roblox?

• Unmatched Scale & Impact: Directly influence how millions of creators build on a platform with 80M+ daily active users.
• Unique Dataset & Domain: Access to a massive, unique corpus of Lua code, creator interactions, and 3D environment context – an incredible sandbox for specializing models.
• Focused Challenge: A chance to dedicate your expertise to solving a critical, concrete problem in AI-assisted creation at the cutting edge.

If you are deeply passionate about pushing the boundaries of LLM capabilities for game development and want to apply your skills to redefine coding assistance for a massive creative community, we encourage you to apply.