About
🔗 GitHub: github.com/gpsyrou
As a Data Scientist, I’m passionate about leveraging…
Activity
2K followers
-
George Spyrou reposted thisGeorge Spyrou reposted thisAnimals and humans get very smart very quickly with vastly smaller amounts of training data than current AI systems. Current LLMs are trained on text data that would take 20,000 years for a human to read. And still, they haven't learned that if A is the same as B, then B is the same as A. Humans get a lot smarter than that with comparatively little training data. Even corvids, parrots, dogs, and octopuses get smarter than that very, very quickly, with only 2 billion neurons and a few trillion "parameters." My money is on new architectures that would learn as efficiently as animals and humans. Using more text data (synthetic or not) is a temporary stopgap made necessary by the limitations of our current approaches. The salvation is in using sensory data, e.g. video, which has higher bandwidth and more internal structure. The total amount of visual data seen by a 2 year-old is larger than the amount of data used to train LLMs, but still pretty reasonable. 2 years = 2x365x12x3600 or roughly 32 million seconds. We have 2 million optical nerve fibers, carrying roughly ten bytes per second each. That's a total of 6E14 bytes. The volume of data for LLM training is typically 1E13 tokens, which is about 2E13 bytes. It's a factor of 30. Importantly, there is more to learn from video than from text because it is more redundant. It tells you a lot about the structure of the world.
-
George Spyrou shared thisThis is by far the best explanation of Gradient Boosting Machines that I have seen out there. The authors have done an amazing job breaking down complex ideas into easy to digest information.This collection of articles provides a full overview of this powerful method, from the math behind the scenes, to how the model learns from data, performs optimisation and combines a set of weak learners to produce a powerful model. #machinelearning #datascience #data #analytics
-
George Spyrou shared thisAfter about 3 months, I have finally completed the Machine Learning Engineer program offered by Udacity. This intensive 3-month course was a great opportunity to leverage cloud technologies like #AWS (#SageMaker, #S3, #Lambda, #APIGateway) to create, train, tune, evaluate and productionize ML models. Over its course, I got the chance to use state of the art algorithms such as XGBoost, LSTM, PCA, and fit them under the general MLOps framework. For my final project, I have chosen to work on one of my favourite fields (NLP), and specifically at the comparison of a "simple" Multinomial Naive Bayes algorithm against the powerful XGBoost for automatic categorization of consumer opinions - which led to some very interesting (and unexpected) results! #machinelearning #cloud #nlp
-
George Spyrou shared thisGeorge Spyrou shared thisU-M is once again ranked the No. 1 public university in the U.S. by The Wall Street Journal and Times Higher Education! myumi.ch/v2gDr #LeadersAndBest #GoBlue #USCollegeRankings
-
George Spyrou shared thisGo blue! #universityofmichiganGeorge Spyrou shared thisFor the sixth straight year, U-M has been named the No. 1 public university in the nation, according to the QS World University Rankings. myumi.ch/jxPBz #LeadersAndBest
-
George Spyrou shared thisGeorge Spyrou shared thisGraduate programs at the University of Michigan remain among the best in the country, according to new rankings from U.S. News & World Report. https://myumi.ch/gjQdr
-
George Spyrou shared thisWhat an amazing article from Ben Denis Shaffer! If you are interested in #AttributionModelling and how you can leverage #MarkovChains on that space, then this is a great introduction for you.George Spyrou shared thisInto to Markov Chain Multi-Touch Attribution by Ben Denis Shaffer
-
George Spyrou shared thisGeorge Spyrou shared thisFor a 10th consecutive year, U-M ranks No.1 in research volume among U.S. public universities, according to a new National Science Foundation (NSF) metrics report. myumi.ch/51rD8
-
George Spyrou liked thisGeorge Spyrou liked thisI've said to not listen to Sam and Dario about the effects of AI on the labor market. They are now saying to not listen to their former selves. Could this have anything to do with their imminent IPOs 🤔 I repeat: listen to economists. https://lnkd.in/eT6939v9
-
George Spyrou liked thisIt was very rewarding working with McKenzie Boutsikakis on this project and throughout the program. Collaborating on a capstone focused on predicting rural hospital closures gave us the opportunity to apply advanced predictive methodologies and machine learning techniques to a meaningful real-world challenge. Through the project, we designed and evaluated multiple supervised learning models using large-scale financial and operational hospital datasets, implemented end-to-end preprocessing and feature engineering pipelines, and explored classification methodologies for identifying early indicators of financial distress and closure risk. The project reinforced the growing role of data-driven decision-making within healthcare operations, particularly in helping organizations identify operational vulnerabilities, improve financial sustainability, and better allocate strategic resources. Working with complex healthcare datasets strengthened our ability to bridge technical analytics with practical business and operational insights, translating model outputs into actionable recommendations for leadership and stakeholders. Proud of everything we accomplished through the University of Michigan - School of Information, and excited to see what’s ahead. Go Blue and Beyond!George Spyrou liked thisExcited to share that I’ve graduated from the University of Michigan - School of Information with a Master of Applied Data Science. Balancing a full-time role while pursuing a rigorous program was challenging, but ultimately a rewarding experience that helped me build and strengthen my skills. As part of the program, I completed my capstone project focused on predicting rural hospital closures using financial and operational data. We built a machine learning pipeline to identify early risk indicators, helping surface insights that could support more proactive operational decision-making. Through this work, I designed and implemented complex machine learning models and developed a deeper ability to translate those outputs into actionable insights, reinforcing my interest in applying data science to real-world problems. Grateful for the experience and looking forward to what’s next! GO BLUE!
-
George Spyrou reacted on thisGeorge Spyrou reacted on thisThrilled and honored to receive the 2026 USF College of Engineering Early Career Research & Entrepreneurship Award! This recognition reflects the dedication of our research group, as well as the support of students and colleagues in the USF Electrical and Computer Engineering Department.
-
George Spyrou liked thisGeorge Spyrou liked thisHonored to have received the Outstanding Undergraduate Teaching Award from the University of South Florida. I'm grateful to my students, colleagues, and USF Electrical and Computer Engineering for their support.
-
George Spyrou liked thisGeorge Spyrou liked thisYou’ll remember this moment forever, Wolverines 🎓 #Victors2026 #MGoGrad
-
George Spyrou reacted on thisGeorge Spyrou reacted on thisIf there is ever an event to restore faith in all that is good about humanity… We were there to cheer on my daughter’s flat mate Sophie who was running to raise funds for Marie Curie UK. Even though Sophie is fit and active and had put in the hours and hours of training required to prepare for the marathon, it was a tough day. Not only because of the miles ahead but she also had an ankle injury and was ill so also was on antibiotics! Despite all that she was absolutely beaming when she ran past us at Tower Bridge which is about halfway. She was still beaming at Mile 22 although when we had a hug she admitted ‘I’m dying’ 😂 Glad to say Sophie finished with a fab time of 4hrs 21mins! Big up to all the runners who took part in the London Marathon yesterday. A wonderful reminder that AI will never over take humans…
-
George Spyrou liked thisGeorge Spyrou liked thisTikTok Shop is growing fast in Spain, and we’re hiring people who want to help build the partnership side of the business. We currently have 2 open roles in Madrid: 👉 TikTok Shop Spain Partnership Manager: https://lnkd.in/gnicFwTG 👉 TikTok Shop Spain Partnership Business Intern: https://lnkd.in/e3NWUggB You’ll be working at the intersection of: - creators / agencies / sellers - content × commerce - globally proven playbooks × local market execution ✅ This is for you if: - You are good at breaking down problems and finding practical solutions - You have strong analytical skills and can use analysis to spot opportunities - You think with structure and frameworks, not just day-to-day tasks - You know how to align with partners and create win-win growth - You care about making the business bigger for everyone involved - You hold yourself to a high professional standard
-
George Spyrou liked thisΔιοργάνωση Οικονομικής Ολυμπιάδας 2026, Τμήμα Οικονομικής Επιστήμης, Πανεπιστήμιο Πειραιώς
Experience
-
-
-
-
-
-
-
Education
-
University of Michigan
-
-
Activities and Societies: Kasper Space Lab
Coursework:
Statistical Computing (Python, C++)
Multivariate Analysis (Machine Learning)
Statistical Learning
Computational Statistics & Algorithms
Statistical Inference
Network Analysis
Applied Bayesian Inference -
-
-
-
-
-
Licenses & Certifications
Volunteer Experience
-
Volunteer
TEDxUniversityofPiraeus
- 2 months
Education
Organizing Committee Member
• Active member of a team of 29 volunteers with a job rotation role
• Immersed in crowd management techniques in order to ensure an effective event
Courses
-
Applied Bayesian Inference
BIOSTAT 682
-
Computational Methods and Tools in Statistics
STATS 506
-
Networks
SI 608
-
Probability and Distribution Theory
STATS 510
-
Real Analysis
ME406
-
Special Topics in Applied Statistics (Data Analysis in Python)
STATS 700
-
Special Topics in Engineering
ENGR 599
-
Statistical Computing
STATS 607
-
Statistical Computing
BIOSTAT 615
-
Statistical Inference
STATS 511
-
Statistical Learning I: Regression
STATS 500
-
Statistical Learning II: Multivariate Analysis
STATS 503
Projects
-
Binary Prediction of Edge Presence Using ERGM's For the International Coffee Trade Network
-
Used Exponential Random Graph Models and Network Analysis techniques in order to analyze the global coffee trade network for three different periods (2005, 2010, 2015).For this purpose we used three types of variables in our analysis: General model variables (name of exporter country, name of importer country etc), Economic/Demographic variables (GDP, population), as well as graph statistics like number of edges, mutual trade relationship between pairs of countries, edge distribution. Our main…
Used Exponential Random Graph Models and Network Analysis techniques in order to analyze the global coffee trade network for three different periods (2005, 2010, 2015).For this purpose we used three types of variables in our analysis: General model variables (name of exporter country, name of importer country etc), Economic/Demographic variables (GDP, population), as well as graph statistics like number of edges, mutual trade relationship between pairs of countries, edge distribution. Our main goal was edge prediction (existence of trade) between countries for different years, along with some Exploratory Data Analysis (EDA).Our final results indicate that wealthy countries have high import-export activity at the coffee trade, while poor countries (which are the lead exporters of coffee) seemed to focus only in exporting (low importing activity).Our final model achieved a prediction accuracy of approximately 86, 5%, which we believe that we can improve by using more complex graph statistics.
For this project we used the Python and R programming languages.Other creators -
-
Classification of Polish Bankruptcy Data via Support Vector Machines,Random Forests and Neural Networks
-
Project used Principal Component Analysis, Neural Networks, Random Forests, and Support Vector Machines in order to classify a dataset of Polish corporate bankruptcy data, finding significant differences in the classification error rates across methods. We followed the whole data analysis procedure: Collect data, Handle missing cases through imputation (Predictive Mean Matching), Dimension Reduction (PCA) and then compare the 3 different Machine Learning Classification methods. We concluded…
Project used Principal Component Analysis, Neural Networks, Random Forests, and Support Vector Machines in order to classify a dataset of Polish corporate bankruptcy data, finding significant differences in the classification error rates across methods. We followed the whole data analysis procedure: Collect data, Handle missing cases through imputation (Predictive Mean Matching), Dimension Reduction (PCA) and then compare the 3 different Machine Learning Classification methods. We concluded that a business analyst, who is more averse on False Negative errors, should choose the SVM method, while an individual who is more averse on False Positive errors, should follow the Random Forests method.
Finally, the project is completed by using the programming language R .Other creators -
Languages
-
Greek
Native or bilingual proficiency
-
English
Native or bilingual proficiency
Other similar profiles
Explore more posts
-
Vincent Kieuvongngam
I’m excited to share my latest blog in the series on context graphs in RDF! In this post, I discuss the ontology discovery process using agent skills. I've also developed an open-source agent skill specifically for this process, with agent first and human-in-the-loop process. In the spirit of the cli vs mcp debate, this skill can be installed via the new GitHub CLI for skill registry. It uses open-source CLI tools for ontology validation and quality checks. You can install it here: gh skill install VincentK1991/present-at-hand ontology-discovery This skill works really well with Claude code and Codex. If you ever have to deal with tedious task of ontology work, why not give this a try; go get yourself a cup of coffee while the agents do the work for you. https://lnkd.in/e27Uh3mV
-
Patrick Li
Check out our latest project funded by AWS! (oss-20b-aci-bench) We're fine-tuning gpt-oss-20b — a model that matches ChatGPT's performance while being lightweight enough to run on a local machine, keeping patient data fully private. The Challenge: - How do you achieve ChatGPT-level performance on a model small enough to run on a MacBook? Why This Matters: - 87% of GPT-4's accuracy - 90x smaller, runs on your macbook - Zero data leaves your laptop Full code & methodology: https://lnkd.in/ev_aXy6B #AWS #OPENAI #HealthcareAI #OpenSource #PatientPrivacy #LLM #ClinicalNLP #HIPAA #DigitalHealth
-
Yedan Zhang, Ph.D.
Excited to share my latest project: a Multi-Agent Forecast Assistant built with LangGraph for demand forecasting Q&A. One recurring challenge in enterprise ML is this: forecast models keep improving, but decision-making doesn’t always scale with them. Business users still struggle to extract decisions from model outputs. Dashboards help, but they rarely answer questions like: • Why was one model selected over another? • How did performance differ by category? • Why did demand spike in Q3? • Which SKUs are most at risk next month? So I built a multi-agent system to bridge that gap and turn forecasts into decision-ready insights. How it works: • A supervisor agent routes natural-language queries to the right specialist • A RAG agent uses hybrid BM25 + semantic retrieval over a forecasting knowledge base • A SQL agent queries model outputs, computes metrics, and generates visualizations • A three-layer guardrail system (input, tool, output) handles PII detection, prompt injection, hallucination filtering, and low-confidence responses Evaluation results: • 4.72 / 5.0 average score across 29 test questions (6 categories) • 100% pass rate • 96% routing accuracy • Latency reduced from 5–6 LLM calls to 2–3 per query What this project demonstrates is that agentic AI can go beyond “chat with your data.” It can make forecasting systems more explainable, more accessible, and more actionable for the people who actually drive decisions. Not a chatbot demo, but a decision-support layer on top of enterprise forecasting systems. GitHub: https://lnkd.in/eCjemweP I’d be glad to connect with others working on agentic AI, RAG, forecasting, and decision-support systems. #AgenticAI #LangGraph #RAG #MachineLearning #Forecasting #GenerativeAI #DataScience #LLM
8 Comments -
Wei Duan
This week, I explored one of Netflix's recent research blogs: 📖 “Recommending for Long-Term Member Satisfaction” 🔗 Read the blog (https://lnkd.in/eKxuyyeB) As a data scientist focused on growing into a Senior DS/ML Engineer role, I’ve been diving deep into real-world ML systems that prioritize long-term value over short-term metrics. This post resonated with me. 🔄 From CTR to Long-Term Value Netflix reframes recommender systems by optimizing for long-term member satisfaction, rather than simply maximizing CTR or immediate engagement. Instead of relying on delayed and noisy signals like retention, they define a proxy reward based on user interactions that are: 1. Timely 2. Sensitive to specific recommendations 3. Aligned with long-term satisfaction 📌 Examples include: 1. Fast season completion 2. Thumbs-down after finishing content 3. Watching just a few minutes 4. Discovering new genres 🤖 Two ML Models, One Goal Offline Delayed Feedback Prediction Models → Estimate delayed or missing signals using historical behavior Online Contextual Bandit Models → Use the proxy rewards to optimize real-time recommendations ⚠️ Remaining Challenge: Online-Offline Metric Disparity Even with proxy rewards, aligning offline training with real-world satisfaction remains tricky. Netflix highlights this as an ongoing research area. 🧠 Open Questions for the Community These questions from the blog sparked deep curiosity for me — and I’d love to hear how others in the DS/ML community think about them: 1. Can we learn a good proxy reward function automatically by correlating behavior with retention? 2. How long should we wait for delayed feedback before using its predicted value in policy training? 3. How can Reinforcement Learning be leveraged to better align policy decisions with long-term member satisfaction? If you’ve tackled similar challenges — in recommender systems, delayed feedback modeling, or contextual bandits — I’d love to hear your thoughts. Let’s learn from each other. 💬 — 🔍 Follow me as I continue learning and sharing from the frontlines of ML, recommender systems, and real-world AI systems. I’m always open to new opportunities to grow, build, and contribute. #DataScience #MachineLearning #RecommenderSystems #MLEngineering #NetflixTech #LongTermSatisfaction #ContextualBandit #RL #DelayedFeedback #CareerGrowth #LearningInPublic #OpenToWork
-
Sam McCormick
Very proud to have worked on developing the first open source modelling suite for Marketing Mix Models (MMM), which we at Mutinex hope will lay the groundwork for a more democratic and transparent MMM landscape. The new validation suite enables practitioners to rigorously test and compare MMM models in a consistent, open framework. It's a step toward greater accountability, better standards, and shared progress in marketing science. We’re excited to share it with the community - contributions and feedback are more than welcome! https://lnkd.in/dmWgKUd2 https://lnkd.in/dWJPtr2m
2 Comments -
Andrii Shchur
🐝 Decentralized AI: why a swarm can beat a giant Big LLMs keep getting bigger—but that’s not always the smartest (or most cost-effective) path. A new approach, Fortytwo, shows how a swarm of diverse models can outperform monoliths while staying private, transparent, and resilient. How it works (in plain terms): 🧠 Many models, two roles: each participant can answer and act as a judge. ⚖️ Pairwise judging + Bradley–Terry ranking: answers are compared head-to-head; a probabilistic model reconstructs a global leaderboard. 🪪 Competence-weighted reputation: votes from proven, consistent judges count more; bad actors get downgraded. ⛓️ On-chain metadata: transparent, tamper-resistant history of behavior and reputation. 🏋️ Proof-of-capability & compute-stake: you earn influence by demonstrating skill on public tasks (e.g., math/coding)—making Sybil attacks economically unattractive. Why this matters: 🔒 Privacy by design (no central point of trust) ⚡ Lower latency + model diversity 🛡️ Byzantine-resistant decision making without expensive ZK proof pipelines Results (as reported): GPQA Diamond: 85.90% vs 68.69% for simple majority (~+25% relative) LiveCodeBench: 84.4% MATH-500: 99.6% AIME 2024 / 2025: 100% / 96.66% Gains persist as the swarm grows—up to ~30 agents—and beat majority voting at all sizes. This is a compelling blueprint for open, robust, and economically balanced AI. Not one super-model, but a network of specialized minds—where truth emerges from many disciplined judgments. #DecentralizedAI #MultiAgent #AIInfra #Blockchain #BradleyTerry #AITrust #RobustAI #Fortytwo #AIResearch
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More