# Spyros Mouselinos Applied AI researcher and ML technical leader in Amsterdam. Multi-agent systems at eBay, video diffusion at Moonvalley, and production machine learning at scale. Portfolio: https://www.mouselinos.com/ Current role: Senior Applied Researcher, Ads at eBay. Based in Amsterdam, Netherlands. Contact: mouselinos.spur.kw@gmail.com CV: https://www.mouselinos.com/spyros-mouselinos-cv.pdf ## Experience ### eBay — Senior Applied Researcher, Ads 2026–Present | Amsterdam, Netherlands Building multi-agent systems from scratch for recommendations, advertising-opportunity discovery, and internal workflow automation. ### Moonvalley — Member of Technical Staff, Foundational AI 2024–2026 | London · Remote Large-scale distributed training with model sharding, tensor and data parallelism; text-to-image/video R&D, preference optimization, and synthetic-data pipelines. ### Synthesa AI — Founding ML Architect 2025–Present · Venture | Remote Designed the LLM platform and multimodal RAG architecture for pharmaceutical and biotech evidence synthesis, using parallel agentic workflows. ### Equinox AI B.V. — Consulting ML Architect Jun–Sep 2024 | The Hague, Netherlands Architected a low-latency trading platform with online inference, reinforcement-learning agents, distributed backtesting, and operational risk controls. ### University of Warsaw — Doctoral Researcher & Teaching Assistant 2021–2026 | Warsaw, Poland Research on reasoning and robustness in language and vision-language models. Taught NLP and visual recognition labs and mentored student research teams. ### DeepSea Technologies — ML Engineer / Technical Lead 2018–2021 | Athens, Greece Led a five-engineer team working on anomaly detection, missing-data modeling, weather-aware routing, and predictive maintenance over sparse and corrupted vessel time series. Served 20M predictions daily. ## Selected work ### Multi-agent AI systems at eBay https://www.mouselinos.com/work/ebay-multi-agent-ai/ Coordinated agents that discover advertising opportunities, make recommendations, and automate internal workflows. Challenge: Turn ambiguous Ads research problems into useful recommendations and automated technical workflows. Contribution: Architected agent roles, orchestration, and implementation from the ground up, owning the path from problem formulation to working systems. Outcome: Delivered agentic systems across distinct eBay Ads use cases, connecting discovery, reasoning, and action. ### Distributed video diffusion training at Moonvalley https://www.mouselinos.com/work/moonvalley-distributed-video-diffusion/ Large-scale distributed model training, alongside research on visual quality and temporal consistency. Challenge: Train large generative models across distributed compute while maintaining visual quality, temporal consistency, and data provenance. Contribution: Built parallel training code with model sharding and tensor and data parallelism. Also led R&D on rendered-text fidelity and motion-aware placement, and developed preference-optimization workflows for 70B-parameter LLM/VLM assistants. Outcome: Trained large models using distributed infrastructure and developed synthetic-data pipelines spanning hundreds of millions of licensed video–text pairs. ### Production machine learning at DeepSea Technologies https://www.mouselinos.com/work/deepsea-production-machine-learning/ Anomaly detection, missing-data modeling, and weather-aware vessel routing across sparse, irregular, and corrupted time series. Challenge: Extract reliable operational signals from missing observations, inconsistent sampling, corrupted sensor data, and changing weather conditions. Contribution: Led a five-engineer ML platform team building anomaly detection, vessel-performance models, predictive maintenance, and weather-aware routing. Worked on missing-data imputation and productionized a multimodal Transformer with monitoring and over-the-air updates. Outcome: 20M predictions daily; 50% less bunker-fuel fraud across 1,000 vessels; routing for 100+ vessels. Additional models reduced fuel consumption by 4% and maintenance costs by 10%. ### Real-time ML architecture at Equinox AI https://www.mouselinos.com/work/equinox-real-time-ml/ Low-latency inference, reinforcement-learning agents, and distributed backtesting for an algorithmic-trading platform. Challenge: Connect real-time order-book and alternative-data feeds to dependable, low-latency model decisions. Contribution: Set the platform architecture, combining online inference, Redis caching, Bayesian-updated RL agents, and distributed Monte Carlo backtesting. Introduced design reviews, runbooks, and Grafana observability. Outcome: Delivered the architecture and operational foundations to the internal team, including statistical risk controls, automated kill switches, and hot-swap rollouts. ### Multimodal RAG and evidence synthesis at Synthesa AI https://www.mouselinos.com/work/synthesa-multimodal-rag/ My venture work: an LLM platform for pharmaceutical and biotech evidence synthesis, with multimodal retrieval and parallel agent workflows. Challenge: Coordinate retrieval and reasoning across scientific documents, tables, and images with predictable, traceable execution. Contribution: Designed the cloud-native LLM platform using Django, Docker, GCP, distributed queues, KV caching, and vLLM. Built parallel language and vision-language agent workflows with vector search, reranking, and policy filters. Outcome: Contributed the ML architecture behind a platform whose published validation screened 270,626 abstracts, with 100% sensitivity, 99.4% specificity, and 91.7% less manual-review workload. ## Research publications - Visual and Language Reasoning in Deep Learning Models (PhD dissertation, 2026). Doctoral research on visual reasoning, code generation, and constructive geometry. https://repozytorium.uw.edu.pl/entities/publication/d6bb00e6-6093-4ca3-bc9c-c01ed5fed0fd - Beyond Lines and Circles: Unveiling the Geometric Reasoning Gap in Large Language Models (EMNLP Findings, 2024). Geometric reasoning, internal dialogue, and role-specialized multi-agent systems. https://aclanthology.org/2024.findings-emnlp.360/ - A Simple, Yet Effective Approach to Finding Biases in Code Generation (ACL Findings, 2023). How function names, specifications, and examples influence generated code. Preprint: 2022. https://aclanthology.org/2023.findings-acl.718/ - Measuring CLEVRness: Black-box Testing of Visual Reasoning Models (ICLR, 2022). Adversarial evaluation of visual reasoning without access to model internals. https://arxiv.org/abs/2202.12162 - MAIN: Multihead-Attention Imputation Networks (IJCNN, 2021). Modeling missingness with attention to improve downstream predictions on incomplete data. https://arxiv.org/abs/2102.05428 - A TensorFlow Extension Framework for Optimized Generation of Hardware CNN Inference Engines (Technologies, 2020). Turning TensorFlow models into configurable hardware inference engines. https://www.mdpi.com/2227-7080/8/1/6 - TF2FPGA: A Framework for Projecting and Accelerating Tensorflow CNNs on FPGA Platforms (IEEE MOCAST, 2019). Mapping and optimizing convolutional neural networks for FPGA acceleration. https://doi.org/10.1109/MOCAST.2019.8741940 ## Skills Research & modeling: Multimodal generative AI · Multi-agent systems · Alignment · Reinforcement learning · RAG · Evaluation & robustness ML systems: PyTorch · TensorFlow · Transformers · vLLM · Ray · ONNX · MLflow · Weights & Biases · Spark · Dask Engineering & infrastructure: Python · SQL · Docker · Kubernetes · GCP · AWS · FastAPI · PostgreSQL · Redis · FAISS · Grafana ## Education PhD, Computer Science — University of Warsaw, 2026. MSc, Data Science — Athens University of Economics and Business, 2021. BSc & MEng, Electrical and Computer Engineering — National Technical University of Athens, 2018. ## Public profiles https://github.com/SpyrosMouselinos https://www.linkedin.com/in/spyridon-mouselinos/ https://scholar.google.com/citations?user=D6TDBuUAAAAJ