Selen Uguroglu, PhD
Research Engineer · RL Environments, Agentic Evaluation & LLM Post-Training
I work on making LLM agents measurably better. My research focuses on building RL environments for training and evaluating tool-using, multi-turn agents (sandboxed tools, user simulators, and benchmarks) and on post-training agents in them with SFT, DPO, and GRPO, using reward signals that hold up beyond the benchmark. Personalization has been the through-line of my career, from recommendations at Netflix to building agents that adapt to the individual people they serve.
I hold a Ph.D. in Language Technologies from Carnegie Mellon University, where I was advised by the late Jaime Carbonell. I'm currently a research / machine learning engineer at Airbnb. Before that I was a research scientist at Netflix and Apple, and co-founded two AI startups.
Research Interests
01RL Environments & Agentic Evaluation
Designing RL environments for tool-using, multi-turn agents: sandboxed tool execution, LLM-based user simulators faithful enough to trust, and benchmarks with LLM-as-a-judge rewards validated against human ratings.
02RL for LLM Post-Training
Reinforcement learning for post-training open-weight models: policy optimization with GRPO, reward design and reward modeling, and guarding against reward hacking, with SFT as a warm start.
03RL for Long-Horizon Agents
Hierarchical reinforcement learning and credit assignment for sparse, delayed rewards in multi-step agent tasks.
04Preference Learning & Off-Policy Evaluation
Contextual bandits, offline policy evaluation, and preference-based training beyond language models, including ranking and personalization.
Earlier work spans contrastive representation learning for recommendation, human-in-the-loop active learning, and query and intent understanding for dialogue systems.
Publications, Service & Talks
-
Workshop Organizer, GRAIL 2026
CIKM 2026. Workshop on Generative, Retrieval-augmented and Agentic Intelligence for Personalization.
-
Invited talks
- 2023Sustainability Solutions with AI, panel speaker
- 2022AI Solutions in Climate, TV guest on Birol Guven's show
- 2021"Your AI or Our AI," Festival of Curiosity
- 2019Netflix PRS Conference
-
Similarity at Netflix: contrastive learning for personalization
NeurIPS Workshop. The self-supervised methods behind "Because You Watched" and "More Like This."
-
Context-Based Endpoint Detection
US Patent 10,186,254 B2 · Apple. A contextual language model for speech endpoint detection in Siri.
-
Robust learning under skewed category distributions
ECML · AAAI. First-author papers on robust and data-efficient learning, transfer, and domain adaptation.
Experience
-
Airbnb
Research / Machine Learning Engineer, AI Platform. LLM agents, evaluation, and post-training.
-
Jori AI, Co-Founder & CEO
Vision-language models for warehouse automation in resale and recycling.
-
Netflix, Research Scientist
Contextual bandits for feed ranking shipped to 200M+ members; contrastive learning framework adopted across production teams.
-
Alkymi, Co-Founder & CTO
Attention-based models for structured data extraction from financial text.
-
Apple, Research Scientist, Siri & Search
Neural query understanding and human-in-the-loop active learning for Siri.
-
Carnegie Mellon University
Language Technologies Institute, School of Computer Science. Richard King Mellon Fellow.
Contact
I'm happy to talk with people working on agent evaluation, post-training, and RL. The best way to reach me is email.