Research Engineer - LLM Post-Training & Agents

Kaon (prev. FlowGPT); San Francisco (CA); Presencial; USD 200.000 - 500.000; Jornada completa

Descripción de la vacante

Kaon, based in the San Francisco Bay Area, is seeking a Research Engineer to advance the models behind our interactive storytelling and character experiences. You will focus on post-training methods to boost model behavior and deploy capable agents with memory, tools, and long-horizon interactions.

In this role, you will build training pipelines (SFT, DPO, GRPO), craft datasets, design reward models, and evaluate improvements through offline tests and live experiments, collaborating with

Formación

• Experience training or post-training large language models, including RL or preference optimization.
• Ability to turn open-ended research problems into clear hypotheses, experiments, and measurable results.
• Strong Python and software engineering fundamentals; C/C++ is a plus.

Responsabilidades

• Build and refine LLM post-training pipelines (SFT, DPO, GRPO).
• Develop training and preference datasets with held-out evaluations.
• Design reward models for narrative quality, memory, and personalization.
• Create agent training environments and evaluations for tool use and memory.
• Collaborate with engineering and product teams to bring improvements into production.
• Conduct offline evaluations and online A/B tests.

Conocimientos

Python

C/C++

Post-training LLMs

RL/Pref optimization

LLM-based agents

Hypothesis testing

Ownership

Herramientas

PyTorch

TensorFlow

Linux/UNIX

Descripción del empleo

Research Engineer – LLM Post-Training & Agents

Kaon (https://www.kaon.io/) builds LLM systems for roleplay, interactive storytelling, and personalized character experiences used by millions of people. We’re looking for a research engineer to improve the models behind these experiences and turn those improvements into capable agents.

Post-training is the core of this role: developing training methods, data, rewards, and evaluations that improve model behavior. You’ll apply that work to agents with strong character consistency, long-term memory, tool use, and coherent multi-turn interactions.

What you’ll work on

• Build and iterate on LLM post-training pipelines, including supervised fine-tuning, preference optimization, and reinforcement learning (SFT, DPO, GRPO, and related methods).
• Develop training and preference datasets from user interactions, with careful data quality controls and held-out evaluations.
• Design reward models and feedback signals for narrative quality, instruction following, character consistency, personalization, and memory accuracy.
• Build agent training environments and evaluations for tool use, memory, and long-horizon interactions.
• Explore memory architectures, context management, and memory consolidation that improve an agent’s behavior over time.
• Run controlled experiments, analyze failures, and validate improvements through offline evaluations and online A/B tests.
• Work with the engineering and product teams to bring research improvements into production.

What we’re looking for

• Strong Python skills and solid software engineering fundamentals; C/C++ experience is a plus.
• Hands‑on experience training or post‑training large language models, including practical familiarity with reinforcement learning or preference optimization.
• Experience building or evaluating LLM‑based agents, or a strong interest backed by relevant projects.
• Ability to turn an open‑ended research problem into a clear hypothesis, an experiment, and a measurable result.
• Ownership from implementation and debugging through evaluation and deployment.

Nice to have

• Work on agent memory, personalized generation, continual learning, reward modeling, or long‑context evaluation.
• Experience with roleplay, virtual characters, interactive storytelling, or conversational AI.
• Research publications or open‑source contributions in relevant areas.

Location: San Francisco Bay Area, on‑site.

Employment: Full‑time.

Compensation: $200,000–$500,000 total compensation (base + equity), depending on experience and impact.