LLM Post-Training Engineer

Single Position

Want to join our team?

View All Jobs

LLM Post-Training Engineer

Yokohama,Yokohama,Japan

+1 more

Apply Now

Use AI to find out how well you match with this job

Upload resume

Job description

Company and benefits

Job ID

787828

Join our Team

About this opportunity

The AI Accelerator within BNEW SAN is on a mission to innovate, incubate, and deliver AI-based solutions that fundamentally redesign Customer Support and Network Deployment — shifting from human-led coordination to composable, agent-centric intelligence. We aim to move beyond simple efficiency gains to create a compounding cycle of optimization, where every interaction generates actionable insights that continuously refine operations, reduce time-to-market, and accelerate the journey toward autonomous network operations.

We are looking for an LLM Post-Training Engineer to own the process of making foundation models genuinely useful in our domain. Pretrained LLMs are powerful starting points — but it is post-training that transforms a general-purpose model into a reliable, domain-aligned agent capable of operating autonomously in the complex world of telecommunications. You will design and execute the training strategies, reward frameworks, and data pipelines that shape how our models reason, respond, and improve — closing the gap between what a model knows out of the box and what our agents need to do in production.

What you will do:

Domain Knowledge Injection via Fine-Tuning

• Design and execute supervised fine-tuning (SFT) pipelines that inject domain-specific knowledge into base LLMs across Customer Support and Network Deployment contexts
• Apply parameter-efficient fine-tuning techniques — including LoRA, QLoRA, and adapter-based methods — to adapt large models effectively within compute and latency constraints
• Collaborate with the AI & Domain Specialist and AI Data & Integration Engineer to ensure fine-tuning datasets are high quality, domain-representative, and correctly formatted for training objectives
• Maintain a structured catalog of fine-tuned model variants, adapter checkpoints, and associated training configurations to support reproducibility and lifecycle governance

Policy-Reward Training & Alignment

• Design and implement policy-reward training pipelines using advanced optimization methods including GRPO (Group Relative Policy Optimization) and SAPO (Soft Adaptive Policy Optimization)
• Define reward models and preference datasets that encode domain-specific quality signals — going beyond generic helpfulness to capture what correct, safe, and operationally appropriate behavior looks like in telco contexts
• Iterate on alignment strategies based on evaluation feedback and production signal — continuously refining model behavior to better match system requirements and domain acceptance criteria
• Stay current with the rapidly evolving post-training research landscape, identifying and assessing new alignment techniques for potential adoption

Automatic Dataset Curation & Synthetic Data Generation

• Build and maintain automated pipelines for domain dataset curation — sourcing, filtering, deduplicating, and formatting data at scale for both SFT and reward training
• Design synthetic data generation strategies that augment real-world datasets — producing diverse, high-quality training examples for scenarios where real data is scarce, sensitive, or imbalanced
• Collaborate with domain experts to validate synthetic data quality and ensure generated examples accurately reflect real-world telco operational scenarios
• Implement data quality gates and automated checks that maintain dataset integrity throughout the curation and generation pipeline

Training Run Management & Experiment Tracking

• Own the end-to-end management of training runs — including infrastructure provisioning, job scheduling, hyperparameter configuration, and run monitoring
• Maintain a rigorous experiment tracking practice, logging all training configurations, dataset versions, evaluation results, and model artifacts to ensure full reproducibility and informed decision-making
• Define and report on key training metrics — loss curves, reward signal trends, evaluation benchmark scores, and alignment quality indicators — providing clear visibility to the AI System Architect and engineering leadership
• Identify and resolve training instabilities, overfitting risks, and performance bottlenecks across the post-training pipeline

You will bring:

• Proven hands-on experience in LLM fine-tuning and post-training, with a track record of adapting large language models for production use cases using SFT, LoRA, or equivalent techniques
• Strong understanding of RLHF, preference optimization, and policy-reward training methods — direct experience with GRPO, DPO, PPO, or similar approaches