AI/ML Research Engineer, LLM Post-Training & Evaluation

About this role

Innodata seeks an AI/ML Research Engineer to design and implement LLM post-training and evaluation pipelines that bridge research and production. You'll work with foundation model builders to develop fine-tuning workflows, evaluation harnesses, and infrastructure for reproducible experimentation at scale.

What you'll do

• Design and build LLM fine-tuning pipelines including SFT, preference optimization, and RLHF workflows
• Implement and optimize evaluation systems for LLMs and multimodal models with offline benchmarks and task-specific harnesses
• Integrate human-in-the-loop and AI-augmented evaluation signals into model development workflows
• Lead technically complex ML projects from customer discussions through implementation and delivery
• Build infrastructure for reproducible experimentation, metrics logging, and regression monitoring
• Diagnose model behavior and pipeline failures, including data issues and evaluation drift

What they're looking for

• Python and production-quality ML code
• PyTorch, JAX, or TensorFlow frameworks
• Hugging Face ecosystem, vLLM, and distributed training stacks
• LLM fine-tuning and post-training techniques (SFT, DPO, RLHF)
• ML evaluation pipeline design and implementation
• Data pipeline and ML systems engineering
• Reproducibility and observability best practices
• Cross-functional collaboration with research and engineering teams

Benefits

• Remote position in United States
• Work with leading AI organizations and foundation model builders
• Contribute to internal R&D and open-source benchmarks
• Mentorship and technical leadership opportunities
• Part of a 36+ year established data engineering company (Nasdaq: INOD)

Likely interview questions

• Walk us through a recent project where you implemented a fine-tuning or preference optimization pipeline—what challenges did you encounter and how did you resolve them?
• How do you approach designing evaluation systems that reliably measure model improvements, and how have you handled evaluation drift or inconsistencies?