Research Engineer Multi-Domain Alignment

IAIRO is seeking a Research Engineer to join our Alignment team. We are committed to advancing the frontier of Sovereign AI by developing compact, "right-sized" models that are as steerable and reliable as their massive counterparts.

In this role, you will bridge the gap between base pretraining and real-world deployment. You wont just be fine-tuning checkpoints; you will be architecting the behavioral logic and safety frameworks for a new class of multimodal SLMs. Your work will focus on multi-domain alignment, ensuring our models can transition seamlessly between specialized fieldssuch as Legal, Healthcare, and Industrial Roboticswhile maintaining rigorous adherence to human intent and cultural values.

Core ResponsibilitiesFrontier Alignment Research: Design and implement scalable alignment pipelines (SFT, DPO, PPO) to optimize 1B7B parameter models for high-stakes, domain-specific tasks.

Advanced Preference Modeling: Architect reward models and preference datasets that capture nuanced domain expertise, moving beyond generic "helpfulness" to expert-level reasoning.

Multi-Domain Synthesis: Develop innovative techniques to mitigate "alignment drift" and "catastrophic forgetting" when models are specialized across disparate industries (e.g., ensuring a model stays factually grounded in domain contexts while remaining flexible in creative ones).

Evaluation & Red-Teaming: Devise rigorous, automated benchmarking suites (LLM-as-a-judge) and adversarial testing frameworks to validate model robustness in "out-of-distribution" scenarios.

Open Source & Transparency: Contribute to the broader AI community by open-sourcing high-quality code and producing reproducible research that impacts the Sovereign AI ecosystem.

Required Skills & ExperienceMasters or PhD in Computer Science, ML, or equivalent practical experience in training large-scale models.

Expertise in Python and PyTorch, specifically within the Hugging Face ecosystem (Transformers, TRL, PEFT, Accelerate).

Proven Alignment Track Record: Significant experience with RLHF (Reinforcement Learning from Human Feedback), Direct Preference Optimization (DPO), and Constitutional AI.

Scaling Knowledge: A deep understanding of Scaling Laws and the "Alignment Tax"knowing how to maximize performance in compute-constrained environments.

Multimodal Familiarity: Experience aligning models that process not just text, but visual and sensor-based data.

Bonus QualificationsPublication Record: Research results published at leading venues such as NeurIPS, ICML, ICLR, or MLSys.

Synthetic Data Engineering: Experience building high-fidelity synthetic data pipelines to improve multi-step reasoning and logic.

Hardware Awareness: Familiarity with optimizing inference engines (vLLM, TensorRT-LLM) or writing custom kernels (Triton/CUDA) for deployment on edge devices. .