Research Engineer, LLM modeling / English

必須

• Master's or Ph.D. in Computer Science, Software Engineering, Artificial Intelligence, Machine Learning, Mathematics, Physics, or related fields
• Experience developing complex agentic systems using LLMs
• Significant hands-on experience in software engineering and ML
• Experience with LLM prompt engineering and/or building products with language models
• Experience with large-scale model training and inference using PyTorch or JAX
• Deep understanding of LLM and Transformer architectures
• Ability to read, reproduce, and improve upon research papers
• Strong implementation skills in Python (production-quality code)

• Language requirement (at least one of the following):

  • Japanese: Fluent — able to discuss product development without friction
  • English: Business level

歓迎

• Experience with large-scale reinforcement learning on language models
• Experience designing and implementing multi-agent systems
• Publications at top-tier conferences (NeurIPS, ICML, ACL, EMNLP, or equivalent)
• Hands-on experience implementing alignment techniques such as RLHF or DPO
• Experience with multimodal models (e.g., Vision-Language models)
• Background in agent evaluation or AI safety research
• Ph.D. in CS, ML, NLP, or a related field
• Ability to communicate research findings in English

使用ツール

Why We're Hiring

• JAPAN AI STUDIO aims to function as "the brain of the enterprise" — integrating every SaaS tool a company uses and enabling AI agents to autonomously execute hundreds of workflows. However, realizing this vision requires breaking through "frontier challenges" that current agent technology cannot solve:
• Reasoning quality limits when searching and integrating information across multiple SaaS platforms
• Long-term memory design that retains context across extended business processes
• Multimodal handling that unifies text, images, audio, and structured data
• Low-latency inference in environments where hundreds of companies operate simultaneously
• Over the past year, adoption of LLM-powered agent systems has accelerated rapidly — from coding and research to customer support and security. Looking ahead to a future where AI agents handle increasingly complex tasks end-to-end or in collaboration with humans, JAPAN AI is strengthening the team that will:
• Build more effective agents for long-horizon tasks
• Design coordination mechanisms for agents to collaborate at various scales and accomplish larger objectives
• Solve the necessary challenges — novel harness design, infrastructure improvements, fine-tuning — to maximize agent performance

Mission

• "Solve the problems that make today's agents give up."
• Take on frontier challenges that today's AI agents cannot solve. Break through the quality limits of reasoning, retrieval/planning, long-term memory, and tool use. Pave the way — through research — for a future where hundreds of workflows running on JAPAN AI STUDIO operate smarter, faster, and more safely.

Role & Expectations

• As a Research Engineer, you will lead cutting-edge and applied research in AI/LLM/ML:
• Conceive, develop, and compare different agent harnesses (memory, context compression, inter-agent communication architectures, etc.)
• Design and implement rigorous quantitative benchmarks for large-scale agentic tasks
• Support automated evaluation of models and prompts, ensuring quality across the entire lifecycle — from training to production
• Collaborate with the product organization to solve the hardest challenges in applying agents to products
• Create and optimize training data mixes to improve agent task performance and usability
• Transfer research outcomes to the Agentic Product Engineer team, raising quality across all products
• Writing papers is not the goal. We prioritize applying research to production and delivering results to users in a live environment serving approximately 200 companies.

Why You'll Love This Role

• _Research that powers "the brain of the enterprise"_ — This is not about improving chatbots. You will build the technical foundation for a next-generation core system where AI autonomously executes operations by integrating all enterprise SaaS — and you will do it through research.
• _Research → Production, directly connected_ — Your methods are immediately deployed to a production environment used by ~200 companies. This is not research that ends with a paper — you will feel real-world impact.
• _Cutting-edge AI research in practice_ — Work on the industry's frontier challenges: breaking reasoning quality limits, designing long-term memory, orchestrating multi-agent coordination, and more.
• _Research and publication, side by side_ — We encourage paper publication and tech blog writing, and actively support collaboration with academic institutions and OSS communities.
• _Impact through technology transfer_ — Transfer your methods to the Agentic Product Engineer team and exercise leadership in raising quality across all products.
• _Rapid-growth environment_ — In a startup that has grown to 200+ people and 9 products in just 3 years, you will have significant autonomy in technical decision-making.

Agent Research & Development

• Conceive, develop, and compare different agent harnesses (memory, context compression, inter-agent communication architectures, etc.)
• Research and develop new reasoning, planning, and retrieval methods
• Develop technologies for multimodal and long-context handling
• Survey, reproduce, and improve upon the latest research papers

Evaluation & Benchmarking

• Design and implement rigorous quantitative benchmarks for large-scale agentic tasks
• Design synthetic data generation and evaluation benchmarks
• Support automated evaluation of models and prompts (across the full lifecycle from training to production)

Production Problem-Solving

• Optimize inference latency and cost (quantization, distillation, caching, etc.)
• Create and optimize training data mixes
• Advance agent evaluation frameworks
• Improve quality and tune performance in production environments

Knowledge Transfer & Outreach

• Transfer technology and mentor the Agentic Product Engineer team
• Collaborate with academic institutions and OSS communities

Key Results (KR/Metrics)

• Benchmark score improvement rate (internal and public benchmarks)
• Number of novel methods shipped to production (per quarter)
• Inference latency and cost reduction rate
• Number of papers and technical blog posts published
• Number of internal knowledge transfers completed

Team Structure

• Approximately 120 members are part of the development organization.
• Research Engineers work across the following groups:
• JAI Lab — AI research and development
• AI & Model — Model training and optimization
• Voice & Tel — Speech AI and telephony systems
• Closely collaborating roles:
• Agentic Product Engineer — Agent feature development (primary research transfer target)
• Agent Harness Engineer — Agent execution infrastructure
• AI Quality Scientist — Evaluation pipeline collaboration
• Product Manager — Product design and prioritization

Tech Stack

• Languages: Python (research / framework), TypeScript / React / Next.js (frontend) / NX
• ML/AI: PyTorch, JAX, Transformers, vLLM, Weights & Biases
• Infrastructure: GCP (containers / K8s), Docker
• Tools: Slack, Confluence, Linear, Google Workspace, GitHub, Notion
• AI Dev Support: Claude Code MAX Plan, Cursor, ChatGPT, Devin
• Hardware: Mac (Apple Silicon), dual monitors