I am a Research Scientist with end-to-end experience in multimodal LLM post-training, automated data curation infrastructure, and on-device deployment. I am currently shipping production data pipelines and LLM-as-a-judge evaluation systems for Ray-Ban Meta AI Glasses, supporting 11+ languages. I hold a Ph.D. in Electrical and Computer Engineering from NYU Tandon School of Engineering, where I was part of the NYU Wireless research center under supervision of Prof. Elza Erkip. My prior research spans machine learning for wireless communications, including deep neural networks, beam alignment, and semantic communications. Comfortable across the stack from training-time research questions to async pipeline architecture and constrained-hardware deployment with two patents, seven publications, and a track record of taking ambiguous problems from prototype to production.

Beyond my professional commitments, I thoroughly enjoy delving into various hobbies. Presently, my passions span a diverse spectrum, encompassing activities such as hiking, climbing, yoga, engaging in film photography, enjoying electronic music, and exploring engaging reads, with a particular focus on science fiction and psychology (current read).

• Bio
• Research
• Get in Touch

Bio

Experience

Full-time

Research Scientist, Multimodal AI — Ray-Ban Meta AI Glasses | September 2025 - Present

• Post-training of multimodal AI models (text, vision, audio) for AR wearables, improving translation quality and expanding multilingual coverage across 11+ target languages
• Architected and shipped a 10-phase automated data curation pipeline that replaced manual workflows, producing high quality training samples across 11 languages
• Built a production translation evaluation and auto-rewrite workflow with iterative improvement loops, raising success rates from 48% to 90% (+40pp) across European languages
• Designed a multi-dimensional LLM-as-a-judge framework from scratch for translation grading, image quality, refusal detection, and helpfulness; calibrated against human annotations and used as a continuous reward signal for RL training
• Migrated synchronous pipeline to a fully async architecture with adaptive concurrency control, buffered checkpoints, and per-phase fault isolation, improving throughput and fault tolerance

Graduate Internship

PhD Software Engineer Intern | 06/2024 - 08/2024

Graduate Intern | 06/2023 - 08/2023

Summer Intern | 06/2022 - 08/2022

Summer Intern | 06/2021 - 08/2021

Undergraduate Internship

Part-time | 08/2019 - 08/2020

Part-time | 10/2018 - 02/2019

Summer Intern | 06/2018 - 07/2018

Summer Intern | 07/2018 - 08/2018

Summer Intern | 06/2017 - 08/2017

Education

Ph.D in Electrical and Computer Engineering | 08/2020 - 08/2025

B.S in Electrical and Electronics Engineering | 08/2015 - 05/2020

Exchange | 02/2019 - 08/2019

Research

Publications

Conference

Hardware Constrained Optimization

• Hardware-Algorithm Co-Design for Interference Mitigation in FR3 Receivers
  Peijie Yang, Ozlem Yildiz, Reza Nikandish, Wenpu Xia, Sundeep Rangan, Elza Erkip, Hamed Rahmani, submitted to Asilomar 2026
  

Frequency Dependent Beamforming

• 3D Beamforming Through Joint Phase-Time Arrays
  Ozlem Yildiz, Ahmad AlAmmouri, Jianhua Mo, Younghan Nam, Elza Erkip, and Jianzhong (Charlie) Zhang, IEEE International Conference on Communications (ICC) 2024.
  

  Abstract: High-frequency wideband cellular communications over mmWave and sub-THz offer the opportunity for high data rates. However, it also presents high path loss, resulting in limited coverage. High-gain beamforming from the antenna array is essential to mitigate the coverage limitations. The conventional phased antenna arrays (PAA) cause high scheduling latency owing to analog beam constraints, i.e., only one frequency-flat beam is generated. Recently introduced joint phase-time array (JPTA) architecture, which utilizes both true-time-delay (TTD) units and phase shifters (PSs), alleviates analog beam constraints by creating multiple frequency-dependent beams for scheduling multiple users at different directions in a frequency-division manner. One class of previous studies offered solutions with "rainbow" beams, which tend to allocate a small bandwidth per beam direction. Another class focused on uniform linear array (ULA) antenna architecture, whose frequency-dependent beams were designed along a single axis of either azimuth or elevation direction. This paper presents a novel 3D beamforming design that maximizes beamforming gain toward desired azimuth and elevation directions and across sub-bands partitioned according to scheduled users’ bandwidth requirements. We provide analytical solutions and iterative algorithms to design the PSs and TTD units for a desired subband beam pattern. Through simulations of the beamforming gain, we observe that our proposed solutions outperform the state-of-the-art solutions reported elsewhere.

Intelligent Reflective Surfaces

• Investigation and Optimization of Secrecy Capacity for Intelligent Reflective Surfaces-Assisted Secure mmWave Indoor Wireless Communication
  Ozlem Yildiz, Mohammad Alavirad, and Tejinder Singh -- IEEE Radio and Wireless Symposium (RWS) 2023.
  

  Slides  

  Abstract: This paper investigates the secrecy performance of an intelligent reflective surface (IRS)-assisted 28 GHz indoor system where the IRS is capable of adjusting the direction and phase shift of reflected signal on its surface and assists a source to communicate with an authenticated user in the presence of a passive eavesdropper. IRSs being promising candidates for millimeter-wave (mmWave) and 6G communication, their deployment has tremendous challenges due to severe information leakage of beam misalignment and costly implementation of continuous phase control. We propose the design of a tile-allocation-and-phase-shift-adjustment strategy for the IRS to optimize the secrecy capacity. In addition to the alignment strategy, the optimal placement to achieve the maximum secrecy capacity is also evaluated through ray-tracing algorithm. Results demonstrate that the proposed IRS-assisted secure algorithm can significantly improve the secrecy capacity performance for an indoor media commutation network.

Beam Alignment

• Hybrid Beam Alignment for Multi-Path Channels: A Group Testing Viewpoint
  Ozlem Yildiz, Abbas Khalili, and Elza Erkip -- IEEE Asilomar Conference on Signals, Systems, and Computers, 2022.
  

  Slides   Poster  

  Abstract: High-frequency bands such as millimeter-wave and terahertz require narrow beams due to path loss and shadowing. Beam alignment (BA) methods allow the transceivers to adjust the directions of these beams efficiently by exploiting the channel sparsity at high frequencies. This paper investigates BA for an uplink scenario, where the channel between the user equipment (UE) and base station (BS) consists of multiple paths. The BS wishes to find the best beams with a given resolution while using the least number of time slots. At each time slot of the BA, the UE transmits a BA packet and the BS uses hybrid beamforming to scan its angular region. To minimize the expected BA duration, a group testing framework is devised, and the associated novel analog and hybrid BA strategies are described. Simulation studies show the performance improvement both in noiseless and realistic 5G mm Wave BA settings.

Journal

Hardware Constrained Optimization

• Hardware Constrained Receiver Under Strong Structured Interference
  Ozlem Yildiz, Sundeep Rangan, Hamed Rahmani, Elza Erkip, submitted to IEEE Transactions on Wireless Communications
  

Federated Learning

• Federated Spatial Reuse Optimization in Next-Generation Decentralized IEEE 802.11 WLANs
  Francesc Wilhelmi, Jernej Hribar, Selim F. Yilmaz, Emre Ozfatura, Kerem Ozfatura, Ozlem Yildiz, Deniz Gündüz, Hao Chen, Xiaoying Ye, Lizhao You, Yulin Shao, Paolo Dini, and Boris Bellalta -- ITU Journal on Future and Evolving Technologies (ITU J-FET) 2022.
  

  Abstract: As wireless standards evolve, more complex functionalities are introduced to address the increasing requirements in terms of throughput, latency, security, and efficiency. To unleash the potential of such new features, artificial intelligence (AI) and machine learning (ML) are currently being exploited for deriving models and protocols from data, rather than by hand-programming. In this paper, we explore the feasibility of applying ML in next-generation wireless local area networks (WLANs). More specifically, we focus on the IEEE 802.11ax spatial reuse (SR) problem and predict its performance through federated learning (FL) models. The set of FL solutions overviewed in this work is part of the 2021 International Telecommunication Union (ITU) AI for 5G Challenge.

Non-Orthogonal Multiple Access

• Power allocation and temporal fair user group scheduling for downlink NOMA
  Eray Erturk, Ozlem Yildiz, Shahram Shahsavari, and Nail Akar -- Telecommun Syst 77, 753–766, 2021.
  

  Abstract: Non-Orthogonal Multiple Access (NOMA) has been proposed as a new radio access technique for cellular networks as an alternative to OMA (Orthogonal Multiple Access) in which the users of a group (pairs or triples of users in a group are considered in this paper) are allowed to use the wireless channel simultaneously. In this paper, for downlink single-input single-output SISO-NOMA, a heuristic power allocation algorithm within a group is first proposed which attempts to ensure that the users of a group benefit from simultaneous transmission equally in terms of achievable throughput. Moreover, a user group scheduling algorithm is proposed for downlink NOMA systems by which a user group is to be dynamically selected for transmission while satisfying long term temporal fairness among the individual contending users. The effectiveness of the proposed power allocation method along with the temporal fair scheduling algorithm for downlink NOMA is validated with simulations and the performance impact of the transmit power and the coverage radius of the base station as well as the number of users are thoroughly studied.

Preprint

Neural Network Learning Capability

• Pathfinding Neural Cellular Automata
  Sam Earle, Ozlem Yildiz, Julian Togelius and Chinmay Hegde
  

  Abstract: Pathfinding makes up an important sub-component of a broad range of complex tasks in AI, such as robot path planning, transport routing, and game playing. While classical algorithms can efficiently compute shortest paths, neural networks could be better suited to adapting these sub-routines to more complex and intractable tasks. As a step toward developing such networks, we hand-code and learn models for Breadth-First Search (BFS), i.e. shortest path finding, using the unified architectural framework of Neural Cellular Automata, which are iterative neural networks with equal-size inputs and outputs. Similarly, we present a neural implementation of Depth-First Search (DFS), and outline how it can be combined with neural BFS to produce an NCA for computing diameter of a graph. We experiment with architectural modifications inspired by these hand-coded NCAs, training networks from scratch to solve the diameter problem on grid mazes while exhibiting strong generalization ability. Finally, we introduce a scheme in which data points are mutated adversarially during training. We find that adversarially evolving mazes leads to increased generalization on out-of-distribution examples, while at the same time generating data-sets with significantly more complex solutions for reasoning tasks.

Patents

Frequency Dependent Beamforming

• 3D Beamforming Through Joint Phase-Time Arrays
  Ozlem Yildiz, Ahmad AlAmmouri, Jianhua Mo, and Younghan Nam -- August 25, 2023, U.S. Patent App. 18/811,643.
  

Intelligent Reflective Surfaces

• Increasing Secrecy Capacity for Intelligent Reflective Surface-Assisted Wireless Communications
  Ozlem Yildiz, Mohammad Alavirad, and Tejinder Singh -- September 18, 2023, U.S. Patent App. 18/469,724.
  

Get in Touch