🛸 Currently mid-mission at Elastix AI · future landings undecided
Mahdyar
Mahdyar
Ravanbakhsh
Principal Research Scientist · LLM Efficiency & Multimodal AI
I build production AI at the intersection of multimodal representation learning, large language models, and model optimization — architect of the neural search powering 50M+ users.
🐈 chief morale officer: Fedrous
01
Profile
AI-native research scientist with 10+ years building production ML systems across multimodal representation learning, large language models, and model optimization. Architect of the next-generation neural search powering 50M+ users. Expert in contrastive learning, vision-language models (CLIP, VLLMs), and LLM fine-tuning for domain-specific applications — with a track record of end-to-end model development, from architecture design through A/B-tested deployment at scale. Now focused on LLM post-training and inference efficiency — quantization and model compression for fast, low-cost deployment.
02
Experience
Research Scientist — LLM Efficiency2025 — Present
Elastix AI
- Research on LLM post-training and inference efficiency — making large models fast and cost-effective to serve in production.
- Quantization and model-compression techniques that cut latency and serving cost while preserving output quality.
Principal Research Scientist — Applied MLFeb 2022 — 2025
Zalando SE · Berlin
- Led production multimodal embedding models (vision + text + color) powering search & recommendations across 25+ European markets for 50M+ customers.
- Architect of the next-gen neural search: multilingual CLIP-based retrieval with strong, A/B-tested engagement lifts across 8 markets.
- Spearheaded an in-house multimodal-LLM initiative: synthetic data pipeline (3M+ items), VLLM fine-tuning (Qwen-VL), secured production roadmap approval.
- Pioneered an LLM-based embedding approach enabling compositional retrieval and cross-category matching; improved non-visual attribute retrieval by 17%.
Postdoctoral Research FellowNov 2019 — Feb 2022
TU Berlin · ERC BigEarth Project
- Built scalable Earth-Observation retrieval with deep metric learning and self-supervised representation learning on multispectral imagery.
- Published on multi-label classification under noisy labels (IEEE TGRS, ICIP, IGARSS).
- Taught graduate courses: Image Processing for Remote Sensing, Computer Vision.
Visiting ResearcherSep 2018 — Mar 2019
SAP AI Research · Berlin
- Human-machine collaborative learning for medical image segmentation (ICASSP 2020 Oral, MIDL 2019).
- Weakly-supervised one-shot segmentation (ICCV-W 2019 Spotlight) — resulted in a US Patent.
03
Selected Projects
Next-Gen Multimodal Embedding Model2025—26
- Led the full lifecycle of the 4th-gen model fusing visual, text, and color modalities.
- Custom RGB embedding distinguishing exact color shades beyond prior versions.
- Coordinated AWS HPC training + large-scale benchmarking for release.
Domain-Adaptive Multimodal LLM2024—25
- Led a domain-tailored VLLM for compositional image-text retrieval.
- Built the pipeline: intent extraction → synthetic queries scaled to 3M+ items.
- Extended to cross-category generalization; secured production greenlight.
Next-Gen Neural Search (Production)2022—24
- Sole architect, prototype → production in 6 months.
- Strong, A/B-tested engagement lifts across 8 markets.
- Became the unified engine for all markets, AI assistants & partners.
Attribute-Conditioned Retrieval2023—24
- Natural-language item editing (“same jacket but in olive”).
- Coordinated large-scale annotation; first in-house labeled dataset for LLM fine-tuning.
04
Selected Publications
2025Y. Deldjoo, N. Rafiee, M. Ravanbakhsh, “Agentic Personalized Fashion Recommendation in the Age of Generative AI,” arXiv.
2022T. Siebert et al., “Multi-modal Fusion Transformer for VQA in Remote Sensing,” SPIE.
2022T. Burgert, M. Ravanbakhsh, B. Demir, “Effects of Label Noise in Multi-label RS Classification,” IEEE TGRS.
2020M. Ravanbakhsh et al., “Human-Machine Collaboration for Medical Image Segmentation,” IEEE ICASSP. 🎙 Oral
2019H. Raza, M. Ravanbakhsh et al., “Weakly Supervised One-Shot Segmentation,” ICCV Workshop. ⭐ Spotlight
2017M. Ravanbakhsh et al., “GANs for Abnormal Event Detection,” IEEE ICIP. 🏅 Microsoft Best Paper
2016H. Mousavi, M. Ravanbakhsh et al., “CNN-aware Binary Map for General Segmentation,” IEEE ICIP. 🏅 Best Paper Finalist
05
Awards & Patents
🏆
Microsoft Best Student Paper — 2nd PlaceIEEE ICIP 2017
🥈
Best Paper Award FinalistIEEE ICIP 2016 — top 7 of ~2,000 submissions
📜
US Patent — Weakly Supervised One-Shot Segmentation190876US01, 2020
🎓
ERC Postdoctoral Research GrantBigEarth Project, TU Berlin, 2019
06
Technical Skills
Core ML
Multimodal Representation LearningContrastive & Self-SupervisedCLIP / Vision-LanguageModel OptimizationGenerative Models (GANs, Diffusion)Few-shot Learning
LLM & Generative AI
LLM Post-Training (SFT, DPO/RLHF)VLLM Fine-tuning (Qwen-VL, LLaMA)RAGSynthetic Data GenerationPrompt Engineering
Efficiency & Optimization
Inference EfficiencyQuantizationModel CompressionKnowledge DistillationModel Optimization
Frameworks
PyTorchTensorFlowHugging FaceOpenCVscikit-learn
Infrastructure
AWS (EC2, S3, HPC/GPU)DockerDistributed TrainingLinuxGit
Research
Scientific ResearchNovel Architecture DesignExperimental DesignTechnical WritingPeer Review & Program Committees
Languages
Python (expert)C / C++MATLAB
07
Education & Leadership
Ph.D., Electrical & Telecommunication Eng.2015—19
University of Genova, Italy
Thesis: Deep Generative Models for Novelty Detection in Data Series
Graduate Research, Computer Science2012—15
University of Ljubljana, Slovenia
B.Eng., Software Engineering2002—07
Sadjad University of Technology, Iran
Leadership & Service
- Technical Lead — cross-functional ML (Product, Engineering, Data Quality).
- Mentorship — junior scientists on ML infra & experimentation.
- Teaching — Lecturer, Image Processing & Computer Vision, TU Berlin (2020—21).
- Program Committee — IEEE IGARSS 2021; reviewer for internal DS conferences.