Konstantinos Nikolakakis

Machine Learning Scientist & Engineer

Machine learning scientist and engineer with a Ph.D. in Electrical and Computer Engineering. At SES AI, I built molecular foundation models and state-of-the-art property-prediction models now in production on the Molecular Universe platform, covering the full ML lifecycle from data curation, large-scale GPU training and benchmarking to cloud deployment. Previously a postdoc at Yale.

πŸ“„ Download CV Get in Touch

Research Interests

🧠

Graph Neural Networks

Learning from graph-structured data and molecular representations

πŸ—οΈ

Foundation Models

Molecular and mixture foundation models for property prediction

⚑

Mixture-of-Experts

Scalable mixture-of-experts models for large, heterogeneous data

πŸš€

Efficient Training

Faster training of deep networks and LLMs, and the theory of why it generalizes

🌐

Federated & Decentralized Learning

Robust federated learning and decentralized ML systems (FEDSTR)

πŸ”¬

AI for Science

Generative and predictive models for materials and electrolyte design

Professional Experience

Machine Learning Scientist

SES AI β€’ Woburn, Massachusetts

Sep 2024 - Oct 2026
  • Modeling: Molecular & mixture foundation models, multi-task learning, (set) transformers, graph neural networks, mixture-of-experts, physics-informed prediction heads, learning to rank
  • Data: Dataset collection & curation, cleaning & preprocessing, feature engineering, large-scale data pipelines
  • Training & Evaluation: Large-scale GPU training (H100/H200, Kubernetes), curriculum learning, ensembles, model & dataset ablation studies, benchmarking, uncertainty & quality estimation
  • Deployment: State-of-the-art molecular property models (melting & boiling point, DFT properties) in production on the Molecular Universe platform
  • Applications: Molecular property prediction, materials discovery, battery electrolytes, large-scale molecular screening

Postdoctoral Associate

Yale University, School of Engineering & Applied Science β€’ New Haven, Connecticut

Jul 2021 - Aug 2024
  • Research Domain: Machine learning theory and algorithms: generalization error analysis, stochastic and zeroth-order optimization, efficient training of deep networks and LLMs, federated and decentralized learning, differential privacy, robustness and risk-aware decision making
  • Publications & Service: Published at NeurIPS, ICLR (2023, 2024), ICASSP and SIAM Journal on Mathematics of Data Science (SIMODS); reviewer for NeurIPS and ICLR
  • Collaborations: Researchers at Yale, ETH Zurich, Delft and UW–Madison
  • Mentoring & Funding: Mentored Ph.D. students; awarded the Anthony Massini Postdoctoral Fellowship

Research Contributions & Method Development

Ph.D. Research (Rutgers) & Postdoctoral Research (Yale)

2016 - 2024
  • Designed and built FEDSTR, a decentralized system for federated learning and LLM training on the Nostr protocol (Rust, Python): job orchestration over Nostr relays, FedAvg aggregation, external model storage, sandboxed execution
  • Introduced (S)GD variants that minimize time to accuracy: 3Γ— faster training for LLMs (GPT-2), 4.3Γ— and 9Γ— faster for ResNet-18 on CIFAR-10 and ImageNet
  • Developed generalization error bounds for SGD, its variants and zeroth-order (derivative-free) learning
  • Designed a robust federated learning algorithm under restricted user availability
  • Established sample-complexity guarantees for quantile multi-armed bandits, and designed the first algorithm for quantile A/B testing under differential privacy
  • Statistical guarantees for learning graphs from noisy data and for prediction on noisy tree-structured models, with applications to recommender systems and epidemic dynamics

Technical Skills

ML Methods & Models

Multi-Task Learning, Fine-Tuning, Mixture-of-Experts, Graph Neural Networks, (Set) Transformers, Representation Learning, Physics-Informed ML, Curriculum Learning, Ensembles, Learning to Rank, Generative Models, Uncertainty Estimation, Federated Learning, Efficient Pretraining of LLMs, A/B Testing, ML System Design, Foundation Models

Tools & Infrastructure

Python, PyTorch, Scikit-learn, XGBoost, RDKit, Snowflake, Docker, Kubernetes, Weights & Biases, Large-Scale GPU Training (H100/H200), MLOps (Model Deployment, Versioning, Batch Inference), FastAPI, C/C++, SQL, Rust

Education

Ph.D. in Electrical and Computer Engineering

Rutgers University β€’ New Brunswick, New Jersey

Feb 2016 - Apr 2021

Thesis: Learning Tree-Structured Models from Noisy Data
GPA: 3.9/4.0

Ph.D. Candidate in Computer Science

Rutgers University β€’ New Brunswick, New Jersey

Sep 2014 - Jan 2016

Focus: Machine Learning and Pattern Recognition
GPA: 3.8/4.0

Diploma in Electrical and Computer Engineering

University of Patras β€’ Patras, Greece

Sep 2009 - Jun 2014

Major: Stochastic Signal Processing and Communications
GPA: 8.3/10 (Top 4% of class)
Award: 1st place Programming Competition (250 participants)

Selected Publications

FEDSTR: Money-In AI-Out | A Decentralized Marketplace for Federated Learning and LLM Training on the NOSTR Protocol

K. Nikolakakis, G. Chantzialexiou, D. Kalogerias β€’ Preprint β€’ arXiv β€’ Code

Select without Fear: Almost All Mini-Batch Schedules Generalize Optimally

K. Nikolakakis, A. Karbasi, D. Kalogerias β€’ SIAM Journal on Mathematics of Data Science (SIMODS) β€’ SIAM

Repeated Random Sampling for Minimizing the Time-to-Accuracy of Learning

P. Okanovic, R. Waleffe, V. Mageirakos, K. Nikolakakis et al. β€’ ICLR 2024 β€’ OpenReview

Federated Learning Under Restricted User Availability

P. Theodoropoulos, K. Nikolakakis, D. Kalogerias β€’ ICASSP 2024 β€’ IEEE Xplore

Beyond Lipschitz: Sharp Generalization and Excess Risk Bounds for Full-Batch Gradient Descent

K. Nikolakakis, F. Haddadpour, A. Karbasi, D. Kalogerias β€’ ICLR 2023 β€’ OpenReview

Black-Box Generalization: Stability of Zeroth-Order Learning

K. Nikolakakis, F. Haddadpour, D. Kalogerias, A. Karbasi β€’ NeurIPS 2022 β€’ Paper

Quantile Multi-Armed Bandits: Optimal Best-Arm Identification and a Differentially Private Scheme

K. Nikolakakis, D. Kalogerias, O. Sheffet, A. Sarwate β€’ IEEE JSAIT 2021 β€’ IEEE Xplore

Predictive Learning on Hidden Tree-Structured Ising Models

K. Nikolakakis, D. Kalogerias, A. Sarwate β€’ JMLR 2021 β€’ JMLR

View All Publications on Google Scholar β†’

Get In Touch

I'm always interested in exciting ML and AI projects, from foundation models and efficient training to production ML systems and AI for science.

πŸ“ Based in Boston, Massachusetts