SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Mirage is an AI-native video platform that uses natural language to orchestrate production and editing, leveraging models that replicate professional editor decisions. The company is backed by top-tier investors including Sequoia, Andreessen Horowitz, Kleiner Perkins, and Index Ventures, and was recently featured in Forbes AI 50 and Fast Company's Most Innovative Companies.
As a Research Engineer focused on Agentic Systems, you will push the boundaries of large language models for multimodal creative tasks. You'll develop novel approaches for building and extending agentic systems that understand and operate over complex, real-world data, particularly video content.
Key responsibilities include:
- Design and build end-to-end agentic systems for creative tasks
- Develop novel approaches for training and adapting large language models that power these agents
- Design new objectives, datasets, and fine-tuning strategies to improve agent behavior and reliability
- Explore multimodal reasoning and structured generation for creative control
- Run systematic experiments to evaluate and improve agent performance in real-world tasks
- Design evaluation frameworks for agentic workflows in video analysis and editing
- Analyze failure modes across the full agent loop (planning, tool use, execution) and iterate on improvements
This is an early-stage opportunity to tackle foundational problems in generative media that remain largely unsolved across the industry. You'll work with an interdisciplinary team addressing some of the most difficult technical and creative challenges in the space.
All roles require in-person presence at the NYC headquarters in Union Square.
QUALIFICATIONS:
- BS/MS/PhD in Computer Science, Machine Learning, or related field
- Strong track record building production ML systems or agentic pipelines
- Deep understanding of transformers and modern LLM techniques
- Experience with fine-tuning, alignment, or post-training methods, especially for adapting models to generate structured outputs or drive tool use
- Comfort owning the full stack, from model-level experiments to deployed agent systems
- Strong experimental rigor and good taste for what makes agents actually work in practice
