2 months ago
Responsibilities
- Own one difficult problem from proposal through system implementation, evaluation, gated production release, and a written retrospective.
- Work in the production codebase with a senior mentor and real production data.
- Develop agent harnesses, inference-time workflows, context engineering, tool and skill designs, orchestration, and self-improvement loops.
- Conduct agent post-training using techniques such as SFT, preference optimization, and reinforcement learning, including reward and supervision design.
- Build training and evaluation environments, verifiable rewards, simulated users, and defenses against reward hacking.
- Create data engines and evaluation systems for production traces, failure mining, labeling, synthetic augmentation, workflow benchmarks, LLM-judge calibration, and reliability measurement.
- Improve model efficiency through routing, ensembles, caching, small-model specialization, serving, distillation, or quantization.
Requirements
- No specific degree is required; strong undergraduates, graduate students, and self-taught builders are welcome.
- Demonstrated depth in machine learning or agent systems.
- Strong ML fundamentals and engineering ability with Python and PyTorch.
- Depth in at least one area such as post-training, reward modeling or LLM judges, agent and tool-use systems, retrieval and memory, or evaluation design.
- Statistical literacy, including the ability to size experiments and interpret results rigorously.
- Preferred qualifications include hands-on post-training with open models using TRL, veRL, OpenRLHF, or a custom loop; interactive agent-environment experience; large-scale trace analysis; data curation or synthetic-data work; serving and efficiency experience; multi-node GPU training; publications or open-source work; and security experience involving prompt injection or data governance.
Benefits
- SF Bay Area, on-site/hybrid work arrangement
- Half-time or full-time for the term
- Flexible start
Categories
AI ResearchML Engineering
About Ema
Ema builds an agentic AI platform that delivers AI employees to automate HR, IT, and Finance workflows across existing enterprise systems. It sells to large organizations via enterprise software and implementation services, offering on-prem and air-gapped deployments and 1,000+ prebuilt integrations. Founded in 2023, it has deployments with Wipro and Hitachi.