Research Engineer

Research Engineer-Applied Research (Agentic AI)

About Pocket FM

Pocket FM is building AI-native products that operate at massive scale across storytelling, content creation, personalization, and user engagement. Our Applied Research team focuses on turning frontier AI research into production systems that solve real-world problems where success is often determined by human judgment, preferences, creativity, and nuanced user outcomes rather than purely verifiable answers.

We're looking for Research Engineers who enjoy working at the intersection of research and product, and who have experience designing, training, evaluating, and improving autonomous AI systems.

What You'll Do

• Design and build agentic AI systems that can reason, plan, use tools, and execute complex multi-step workflows.
• Develop domain-specific agent architectures tailored to challenging product problems rather than generic benchmark tasks.
• Formulate real-world problems as sequential decision-making problems and identify effective learning signals for improvement. Agentic systems increasingly benefit from reinforcement learning and environment-driven optimization approaches that align agent behavior with domain objectives.
• Create simulated or real-world environments that allow agents to learn, explore, and improve on metrics that matter to users and the business.
• Build evaluation frameworks, benchmarks, and measurement systems that capture progress in complex, subjective, and human-centered domains.
• Design reward models, preference-learning systems, or other feedback mechanisms where objective ground truth may not exist.
• Run rigorous experimentation to understand agent behavior, failure modes, and opportunities for improvement.
• Collaborate closely with product, engineering, content, and data teams to translate research into production impact.
• Stay current with emerging research in reinforcement learning, agentic systems, reasoning, multimodal AI, and generative models.

What We're Looking For

Strongly Preferred

• Experience building and deploying AI agents or autonomous systems in production environments.
• Demonstrated ability to adapt agent architectures to domain-specific problems rather than simply applying off-the-shelf frameworks.
• Experience with reinforcement learning, preference optimization, online learning, or other approaches that improve agent performance through interaction and feedback. Modern agentic systems increasingly leverage RL-style training and environment interaction to optimize long-horizon behavior.
• Experience designing environments, simulators, synthetic users, or feedback loops that enable continuous agent improvement.
• Strong understanding of evaluation methodologies for LLMs and agents.
• Experience creating benchmarks, test suites, rubrics, or evaluation frameworks for complex domains.
• Ability to reason about subjective metrics such as:
• User satisfaction
• Content quality
• Creativity
• Helpfulness
• Engagement
• Preference alignment
• Trust and safety
• Strong software engineering fundamentals with proficiency in Python and modern ML tooling.

Particularly Exciting Backgrounds

We are especially interested in candidates who have worked in domains where:

• Success cannot be determined by a single correct answer.
• Human judgment plays a central role in evaluation.
• User preferences are diverse and evolving.
• Evaluation requires nuanced qualitative assessment.
• Systems must optimize for long-term outcomes rather than short-term accuracy.

Examples include:

• Consumer AI products
• Creative tools
• Recommendation systems
• Conversational AI
• Content generation
• Education
• Gaming
• Human-AI collaboration systems
• Social or community platforms

Bonus Points

• Experience with diffusion models, image generation, video generation, or multimodal AI.
• Research experience in RLHF, RLAIF, preference learning, reward modeling, or agentic RL.
• Experience with multi-agent systems.
• Publications at top-tier ML conferences.
• Experience building large-scale evaluation infrastructure.
• Familiarity with synthetic data generation and simulation-based training.
• Experience deploying AI systems that learn from real user interactions.
• You're an active consumer of comics, anime, manga, webtoons, or related storytelling formats. We believe that people who deeply enjoy these mediums often bring stronger intuition when designing, evaluating, and improving AI systems that interact with creative content and audience preferences.

What Success Looks Like

Within this role, you'll help answer questions such as:

• How do we teach agents to improve on metrics that humans care about but are difficult to measure directly
• How do we build evaluation systems for creativity, quality, and user satisfaction
• How do we create environments where agents can learn safely and efficiently
• How do we align autonomous systems with nuanced product goals
• How do we move beyond benchmark performance and optimize for real-world impact

If you enjoy building agents, designing feedback loops, creating evaluation systems, and tackling messy real-world problems where human judgment matters as much as model capability, we'd love to talk.