Irina Lebedeva, PhD

• Fine-tuned LLMs with LoRA for automotive requirements engineering, improving model accuracy by 30%.
• Achieved accuracy gains through synthetic data generation, targeted dataset refinement, hyperparameter tuning, and evaluation-driven iteration.
• Developed robust training and evaluation pipelines to identify model weaknesses and improve domain-specific performance.

• Built and deployed a lightweight multi-label image-tagging model trained on synthetic VLM/LLM labels.
• Fine-tuned VLMs with Megatron-DeepSpeed (tensor + pipeline parallelism, ZeRO-3) for large-scale tagging and captioning.
• Designed a low-shot, identity-preserving image generator (LCM-SDXL + IP-Adapter + ControlNet).
• Developed text-to-video pipelines using WAN-based models and LoRA.
• Shipped a face age-estimation service (fine-tuned AuraFace, served via FastAPI).

While primarily an engineering role — from system design to deployment and production — it also involved research: successfully defended postdoctoral research on computer vision, hyperbolic deep learning, and interpretable AI, and published peer-reviewed papers.

Engineering work included, but was not limited to:

• Fine-tuned generative diffusion models using Stable Diffusion and LoRA for conditional face synthesis with improved cultural and social diversity.
• Built and deployed an AI onboarding assistant using LangChain, GPT-4, RAG, ChromaDB, and FastAPI, and fine-tuned LLaMA for paper summarization and QA.

• Built VAE- and GAN-based models for image editing, makeup transfer, and masked-face inpainting.
• Shipped real-time emotion- and face-recognition systems with PyTorch, OpenCV, and Docker.
• Delivered forecasting models with XGBoost and scikit-learn; advanced facial-attractiveness research and released a labeled dataset.

• Owned the full software development lifecycle, designed SQL databases, and mentored engineering teams.