Staff ML Engineer: LLM Training & Data Pipelines

Belvedir (Fractal Machine Research, Inc.); San Francisco (CA); On-site; USD 150,000 - 230,000; Full time

Job summary

Belvedir (Fractal Machine Research, Inc.) is seeking a hands-on ML engineer to design and run training experiments across SFT, distillation, RL, and RLHF, working directly with customer teams to understand objectives and deliver practical results.

You will build pipelines to collect, clean, and synthesize data, create robust evaluations, and turn wins into repeatable training recipes and tools for the team, while collaborating closely with founders.

Qualifications

• Experience with LLM training: post-training, RL, RLHF, evals, or synthetic data.
• You have trained or fine-tuned models on real, messy data.
• Know what makes an eval trustworthy, and how models game one that is not.
• You can work directly with customers and translate what you hear into what to train next.
• Founders or early-engineer experience is a plus.

Responsibilities

• Design and run training experiments: SFT, distillation, RL on customer tasks.
• Build pipelines that turn raw customer data into high-quality training sets: collection, cleaning, filtering, and synthesis.
• Build task-specific evaluations and use them to drive decisions about data and training recipes.
• Turn one-off customer wins into repeatable training recipes and tooling the whole team can use.
• Work directly with customers and founders to decide what to train next.

Skills

LLM training

SFT

Distillation

RLHF

Evaluation design

Data cleaning & synthesis

Customer collaboration

End-to-end ownership

Job description

Belvedir (Fractal Machine Research, Inc.) is seeking a hands-on ML engineer to design and run training experiments across SFT, distillation, RL, and RLHF, working directly with customer teams to understand objectives and deliver practical results.

You will build pipelines to collect, clean, and synthesize data, create robust evaluations, and turn wins into repeatable training recipes and tools for the team, while collaborating closely with founders.