About the job
At Sciforium, we are pioneering the future of AI infrastructure by creating cutting-edge multimodal AI models and a proprietary, high-efficiency serving platform. With substantial financial backing and direct support from AMD engineers, our team is rapidly expanding as we develop the comprehensive stack that drives advanced AI models and real-time applications.
About the Role
In the capacity of a Machine Learning Engineer, you will engage with the entire foundation-model stack, encompassing pretraining and scaling, post-training and Reinforcement Learning, sandbox environments for evaluation and agentic learning, and deployment + inference optimization. You’ll have the opportunity to rapidly iterate on research ideas, contribute to production-grade infrastructure, and help deliver models capable of addressing real-world challenges at scale.
Your Responsibilities
This position offers diverse tracks - candidates can specialize or contribute across multiple areas. Key responsibilities include:
Pretraining & Scaling
- Train expansive byte-native foundation models utilizing vast, heterogeneous data sources.
- Formulate stable training methodologies and scaling laws tailored for innovative architectures.
- Enhance throughput, memory efficiency, and resource utilization across extensive GPU clusters.
- Establish and maintain distributed training infrastructures alongside fault-tolerant pipelines.
Post-training & Reinforcement Learning
- Build out post-training frameworks (SFT, preference optimization, RLHF/RLAIF, RL).
- Curate and produce specialized datasets aimed at enhancing specific model capabilities.
- Develop reward models and evaluation systems to facilitate ongoing improvements.
- Investigate inference-time learning and computational strategies to boost performance.
Sandbox Environments & Evaluation
- Create scalable sandbox environments for agent assessment and learning.
- Generate realistic, high-signal automated evaluations for reasoning, tool usage, and safety.
- Design both offline and online environments that support RL-style training at scale.
- Implement instrumentation for observability, reproducibility, and rapid iteration.
Deployment & Inference Optimization
- Optimize deployment strategies to ensure models are efficient and effective in real-world applications.
