About
Senior AI Research Engineer with 8+ years of experience across LLMs, multimodal and generative AI, computer vision, and production ML. PhD in Computer Science with peer-reviewed Q1 publications and 130+ Google Scholar citations. Experienced in bridging research and engineering. Background includes work at AI startups, research labs, and industry teams, as well as earlier software engineering experience with team leadership.
Skills
LLMs & VLMs: multimodal learning, reasoning models, fine-tuning, test-time compute
Reinforcement learning & post-training: SFT, DPO, ORPO, RLHF, GRPO, reward modeling
Generative AI: diffusion models (SDXL, FLUX, Wan), GANs, VAEs, image and video generation
Distributed training: DeepSpeed (ZeRO-3), Megatron, FSDP
Model & GPU optimization: distillation, quantization, inference optimization, serving (vLLM, SGLang, NVIDIA Triton)
Data & evaluation: synthetic data generation, LLM evaluation, LLM-as-a-judge
Experience
- Post-trained Qwen LLMs and VLMs (SFT, LoRA, DPO, ORPO) to detect and explain quality issues in automotive requirements, with a ~30% relative improvement.
- Improved model reasoning with reinforcement learning with verifiable rewards (GRPO, RLVR), synthetic data generation and LLM-as-a-judge evaluation.
- Optimized and served models in production with quantization, GPU profiling, vLLM and SGLang; redesigned OCR and document processing for 5× faster parsing.
- Fine-tuned a vision-language model for image and video captioning and video understanding, with distributed training (Megatron, DeepSpeed ZeRO-3, FSDP, mixed precision).
- Built few-shot identity-preserving image generation and text-to-video models (SDXL, FLUX, ControlNet, IP-Adapter, Wan, LoRA).
- Distilled VLM-generated synthetic labels into a lightweight multimodal tagger for content search and recommendation; with quantization and NVIDIA Triton, cut latency by 74%.
- Designed hyperbolic neural networks and transformer architectures that embed visual features in hyperbolic space, improving the interpretability of learned representations.
- Developed aesthetic preference learning (human preference modeling): user-adaptive multimodal models that learn an individual’s taste from a few ratings.
- Fine-tuned diffusion models (LoRA) to generate images that match a user’s aesthetic preferences.
- Completed the postdoctoral program with a successful final defense; published the results in peer-reviewed journals.
- Advanced the state of the art in personalized facial attractiveness and aesthetic preference prediction with few-shot and meta-learning; published in Q1 journals.
- Created and publicly released MEBeauty, a multi-ethnic dataset and benchmark with human preference ratings; developed visual-embedding recommender models for the cold-start problem.
- Built generative models (diffusion, GAN, VAE) for face editing, image enhancement and inpainting in PyTorch.
- Built real-time face analysis systems (face recognition, age and emotion estimation) with PyTorch and OpenCV.
- Supervised and mentored developers while managing the complete software development lifecycle.
- Software development and maintenance; SQL database design and administration.
Education
PhD, Computer Science — East China University of Science and Technology, 2017–2022
Outstanding International Graduate (2022)
MSc, Computer Science — Shanghai Jiao Tong University, 2013–2017
Excellent International Student (2015)
BEng, Computer Science & Engineering — National Research Nuclear University MEPhI, Russia, 2005–2009
Selected Publications
Full list on Google Scholar.
Peer Review
Conferences: NeurIPS 2026 (Ethics Review) · ECCV 2026 (WiCV Workshop)
Journals: Neural Computing and Applications · Computer Vision and Image Understanding
Languages
English (professional) · Russian (native) · Chinese (intermediate) · German (basic)
Open Science
A multi-ethnic facial beauty dataset collected in the wild, with human attractiveness ratings, first published in 2022 and now fully open and reproducible.
- Dataset: about 2,500 faces from six ethnic groups, of different genders and ages, cleaned, with person-level splits and an improved label.
- Benchmark: 21 methods, from 2006 to today's foundation models, including my own approach, all under one protocol.
- Reference model: a DINOv2-based model, ready to use in two lines of Python, with its accuracy across ethnic groups reported openly.
- Publication: the original peer-reviewed paper, "MEBeauty: A Multi-Ethnic Facial Beauty Dataset in the Wild", Neural Computing & Applications, 2022.
For non-commercial academic research only.
News
- Oct 2026Released the updated MEBeauty dataset on Hugging Face, together with an open benchmark of 21 methods and a reference DINOv2 model.
- 2026Serving as a peer reviewer for NeurIPS 2026 (ethics review) and the WiCV workshop at ECCV 2026.
- Mar 2026Joined Raiqon as a Machine Learning Engineer, working on LLM and VLM post-training and reinforcement learning.
- 2025Paper published in the Journal of Visual Communication and Image Representation: GAN Semantics for Personalized Facial Beauty Synthesis and Enhancement.
Contact
Open to research and engineering roles in LLM post-training, multimodal AI, and efficient AI, and to research collaborations.
dr.irina.lebedeva@gmail.com