Original listing text, shown exactly as published by the company.
Responsibilities
- Architect and orchestrate massively parallel AI research workflows — design experiments that leverage fleets of agentic AI systems to explore hypothesis spaces, hyperparameter landscapes, and architectural variations at a scale and speed no single researcher could achieve alone
- Design, train, and iterate on models across the full GenAI stack — LLMs, VLMs, embedding models, rerankers, and reward models — using agentic pipelines that autonomously manage data preprocessing, training runs, evaluation sweeps, and result synthesis
- Go deep: push the frontier of domain-specific AI — conduct rigorous, first-principles research into model architectures, training dynamics, reinforcement learning, and knowledge representation, using AI agents to accelerate literature review, ablation studies, and mathematical analysis
- Go broad: span disciplines and modalities — amplify your expertise across NLP, computer vision, multimodal understanding, agentic reasoning, and domain science by delegating exploration, prototyping, and benchmarking to parallel agent systems so you can synthesize insights across fields simultaneously
- Build agentic research infrastructure — develop and contribute to shared tooling, libraries, and platforms that enable every researcher on the team to orchestrate autonomous experiment pipelines, data processing workflows, and evaluation harnesses at scale
- Ship research into production at velocity — collaborate with engineering, product, and domain experts to integrate breakthroughs into the platform rapidly, using agentic CI/CD and automated integration testing to compress the research-to-deployment cycle
- Amplify collective intelligence — document findings, publish at top-tier venues, and build internal knowledge systems that agentic tools can index and reason over — turning every insight into a force multiplier for the entire team
- Continuously raise the ceiling on human potential — proactively identify bottlenecks in your own workflow and the team's, then design or adopt efficient, scalable solutions that eliminate them — treating your own augmentation as a core research output
Required Qualifications
- Education: PhD in Computer Science, Machine Learning, or a related field; or MSc with 4+ years of post-graduation research experience.
- Model development: You have trained or fine-tuned at least one neural model end-to-end — data preparation through evaluation. You understand why your model converges or doesn't, not just how to launch a training run.
- Technical foundations: Strong working knowledge of probability, optimization, and linear algebra applied to at least one of: NLP, computer vision, reinforcement learning, or information retrieval. You can derive the math behind the methods you use.
- Infrastructure: Experience building training or evaluation pipelines that handle real data — preprocessing, distributed computation, experiment tracking, and reproducibility.
- Software engineering: Production-quality Python. You write code others can read, test, and extend. Fluent with Git and collaborative development workflows.
Preferred Qualifications
- Experience with distributed training frameworks (PyTorch DDP, DeepSpeed, FSDP) — you understand gradient synchronization and can debug multi-GPU failures.
- Published at NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, or equivalent. Quality of contribution matters more than count.
- Hands-on experience with post-training methods (RLHF, DPO, reward modeling) — beyond reading papers.
- Practical cloud infrastructure experience (AWS, GCP, or Azure) for ML workloads — you can provision resources, manage jobs, and troubleshoot training failures.
Professional Attributes (Code42)
- Practice Humility: You ask questions even when you think you know the answer. You seek feedback early, learn from anyone regardless of title, and treat every experiment — especially the failures — as data.
- Bias for Outcomes: You measure your work by what changed, not what you tried. You ship results, not slide decks. When a deadline is real, you find a way.
- Care Deeply: You treat every problem as yours to solve. You review your own work with the rigor you'd want from a reviewer. You help teammates without being asked.
- Dare to Do the Impossible & Embrace Scarcity: You set goals that make you uncomfortable. When told something can't be done, you find a way or a better question. Constraints sharpen your thinking, not slow it down.
- Build a Better World: You believe AI should make things meaningfully better for real people. You hold yourself accountable not just for whether your model works, but for what it does in the world.
If you're ready to join a team that's changing the game, apply now to become a part of the Articul8 team.
