sm✳

APPLIED MACHINE LEARNING RESEARCHER & ENGINEER

Spyros
Mouselinos.

Applied machine learning researcher and engineer with eight years across production ML and research. I build recommendations and Ads systems at eBay, with earlier work in multimodal foundation models and ML platforms.

Explore my work
Amsterdam, NL / Currently building at eBay

Short on time? Start with the essentials.

+10% CTR+5% GMV on the tested surface
eBay mobile recommendations
70B rewriterSFT + GRPO, served with vLLM
Moonvalley public product
20M / dayVessel-monitoring predictions
DeepSea Technologies
ICLR · ACL · EMNLPFirst-author research
Reasoning & robustness

01 / SELECTED WORK

Research instincts.
Real-world outcomes.

Different domains. The same thread:
understand the problem, build the system,
make it count.

RECOMMENDATIONS & ADS01

eBay / Jan 2026–Present

From signals to measurable lift

A complementary-product recommender, reusable agentic analytics, and an eBay Live pacing simulator.

Contextual banditsRecommender systemsAgentic analytics
Inside the work

The challenge

Extend similar-item recommendations to complementary products, make fragmented Ads investigations faster, and compare budget-allocation policies under uncertain live-auction conditions.

My contribution

Jointly designed a mobile recommender using offline LLM-derived intent profiles, eBay embeddings, and a contextual bandit before retrieval and reranking. Originated the analytics framework with read-only SQL, permission, and PII checks; wrote its core, reviewed integration by three engineers, and coordinated deployment. Also built the eBay Live pacing simulator.

The outcome

The recommender's online A/B test delivered relative lifts of 10% CTR and 5% GMV on the tested surface. Other Ads teams forked and repurposed the analytics framework; users reported one representative investigation falling from 10 to 3 hours.

Read the eBay work story ↗
MULTIMODAL FOUNDATION MODELS02

Moonvalley / Sep 2024–Jan 2026

Making video models follow intent

Text conditioning, a 70B prompt rewriter, and dataset pipelines for customer-facing video generation.

Multimodal conditioningLLM post-trainingData curation
Inside the work

The challenge

Improve rendered text and branding while closing the mismatch between short user requests and the detailed captions used to train video models.

My contribution

Tested adapters against conditioning integrated into from-scratch training and co-developed multi-encoder text conditioning, text-to-render extraction, and targeted attention weighting. Owned a 70B Llama-family prompt rewriter from hundreds of thousands of training pairs through SFT, GRPO, and vLLM serving. Built a pipeline that annotated hundreds of millions of licensed videos.

The outcome

The conditioning approach shipped in the customer-facing model, the prompt rewriter entered the public generation product, and the video labels shaped a later production training curriculum while helping avoid unsuitable supplier purchases worth hundreds of thousands of dollars.

Read the Moonvalley work story ↗
PRODUCTION ML03

DeepSea Technologies / Apr 2018–Mar 2021

Taking maritime ML into production

A multimodal Transformer for vessel anomaly detection, backed by repeatable experiment and release workflows.

Anomaly detectionRay ServeML platform
Inside the work

The challenge

Run vessel-monitoring inference at production volume while giving a five-engineer team dependable, shared experiment and release practices.

My contribution

Productionized a PyTorch multimodal Transformer with Ray Serve. Led ML platform engineering for a five-engineer team, establishing MLflow experiment tracking, Jenkins CI/CD, blue-green deployment, and over-the-air updates.

The outcome

The model served approximately 20 million predictions a day in a vessel-monitoring system, supported by reusable delivery workflows for the wider ML team.

Read the DeepSea Technologies work story ↗
REAL-TIME ML SYSTEMS04

Equinox AI B.V. / Jun–Sep 2024

From market data to controlled execution

Low-latency inference, reinforcement-learning agents, and distributed backtesting for an algorithmic-trading platform.

Online inferenceReinforcement learningRisk controls
Inside the work

The challenge

Connect real-time order-book and alternative-data feeds to dependable, low-latency model decisions.

My 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.

The outcome

Delivered the architecture and operational foundations to the internal team, including statistical risk controls, automated kill switches, and hot-swap rollouts.

Read the Equinox AI B.V. work story ↗
SCIENTIFIC EVIDENCE SYNTHESIS05

Synthesa AI / Feb 2025–Present · Venture

Research evidence, made operational

An evidence-synthesis platform used in paid workflows by pharmacologists and biomedical researchers.

Hybrid retrievalRerankingProduction inference
Inside the work

The challenge

Support systematic reviews and evidence reports spanning papers, trial records, and private research materials—sometimes up to 1,000 source documents.

My contribution

Built and deployed the platform, then owned retrieval and production inference for compute-heavy research jobs alongside engineering and pharmacology co-founders.

The outcome

The platform is used in paid research workflows, with outputs evaluated through expert review and meta-analysis projects.

Read the Synthesa AI work story ↗

02 / EXPERIENCE

From the lab.
Into the world.

Research, engineering, and technical leadership — with ownership all the way through.

Netherlands01 / 06
eBay — Amsterdam
eBayAmsterdam
52.37° N
4.90° E

eBay

Jan 2026–Present

Senior Applied Researcher, Ads

Shipped a mobile complementary-product recommender with a contextual bandit, originated reusable agentic analytics adopted by other Ads teams, and built an eBay Live pacing simulator.

Amsterdam, Netherlands

Moonvalley

Sep 2024–Jan 2026

Member of Technical Staff, Foundational AI

Developed multimodal conditioning for a customer-facing video model, owned a production 70B prompt rewriter, and annotated hundreds of millions of licensed videos for training-data curation.

London · Remote

Synthesa AI

Feb 2025–Present · Venture

ML Architect

Built and deployed an evidence-synthesis platform used in paid research workflows, owning retrieval and production inference for jobs spanning up to 1,000 source documents.

Independent venture · concurrent with employment

Equinox AI B.V.

Jun–Sep 2024

Consulting ML Architect

Architected a low-latency trading platform with online inference, reinforcement-learning agents, distributed backtesting, and operational risk controls.

The Hague, Netherlands

University of Warsaw

Sep 2021–Dec 2025

Doctoral Researcher & Teaching Assistant

Developed adversarial evaluation and multi-agent reasoning methods across visual question answering, code generation, and geometry; mentored a four-student vision-language research team.

Warsaw, Poland

DeepSea Technologies

Apr 2018–Mar 2021

ML Engineer

Productionized a PyTorch multimodal Transformer serving approximately 20 million predictions a day and led ML platform engineering for a five-engineer team.

Athens, Greece

03 / RESEARCH & IDEAS

But does it
actually reason?

A recurring question in my work: where does
pattern matching end and reasoning begin?
All publications on Google Scholar ↗
Read the research notes ↗

04 / THE HUMAN IN THE LOOP

Hi again.
I’m Spyros.

Spyros Mouselinos’s GitHub profile image

Based in Amsterdam.
Greek roots. A curious mind.

I like the space between “what if?” and “it works.”

My work combines applied machine learning research, engineering, and technical leadership. I’ve shipped recommendation systems, foundation-model components, and production ML infrastructure; led platform work for a five-engineer team; and taught master’s-level NLP and visual recognition.

At the University of Warsaw, my doctoral research focused on visual and language reasoning in deep learning models. I developed adversarial evaluations and multi-agent methods to separate reasoning from shortcuts— a question that still shapes how I build.

Education

PhD, Computer Science, with honorsUniversity of Warsaw · 2026

MSc, Data ScienceAthens University of Economics and Business · 2021

BSc & MEng, Electrical and Computer EngineeringNational Technical University of Athens · 2018

Modeling & evaluation

Recommendation systems · Contextual bandits · Multimodal diffusion models · LLM post-training (SFT, GRPO) · Model evaluation · A/B testing

Retrieval & agents

Hybrid retrieval · Reranking · Multi-agent orchestration

Engineering

Python · SQL · PyTorch · vLLM · Ray · GCP · Docker · Kubernetes

Greek Native/ English C2

THE 30-SECOND OVERVIEW

Spyros Mouselinos

Applied Machine Learning Researcher & Engineer · Amsterdam

Eight years across production ML and research, spanning recommendations, personalization, multimodal foundation models, and reusable ML systems.

Now
Senior Applied Researcher, Ads at eBay — shipped a mobile recommender with +10% CTR and +5% GMV on the tested surface, and built reusable agentic analytics adopted by other Ads teams.
Previously
Multimodal conditioning and a production 70B prompt rewriter at Moonvalley; ML platform leadership at DeepSea, serving approximately 20 million predictions a day.
Research
PhD in Computer Science with honors, University of Warsaw (2026). First author at ICLR and Findings of ACL and EMNLP.
Core skills
Recommendation systems, contextual bandits, multimodal diffusion, LLM post-training, hybrid retrieval, multi-agent orchestration, and PyTorch.
Get in touch ↗
Connect on LinkedIn ↗