IASON SOMOGLOU
ABOUT THE PROGRAM
Greetings, User. I am an AI developer and Machine Learning engineer currently pursuing a BSc in Artificial Intelligence at Vrije Universiteit Amsterdam, with a Minor in Deep Learning.
My programs have been running in the real world since 2022 — at Tesla, SkillLab.io, EY, and Martech Tribe — solving complex problems in NLP, ML pipelines, transformer architectures, and intelligent systems.
I was recognised by Forbes 30 Under 30 Greece, hold a patent for an assistive cane, and came 1st place both nationally and globally at the Robotics Competitions — some of the world's largest global robotics competitions (WRO,FGC, etc.).
COMBAT HISTORY
TESLA
MAR 2025 – NOV 2025Deployed end-to-end NLP/ML pipeline for vehicle diagnostics. Trained transformer-based models and sentence embeddings for semantic search and intent classification. Containerised services with FastAPI & Docker.
SKILLLAB.IO
APR 2024 – MAR 2025Developed replication recommender systems using BERT models and cosine similarity. Identified bugs in production systems, upgraded libraries, and streamlined CI/CD with GitHub and Jira.
EY (ERNST & YOUNG)
JUN 2023 – AUG 2023Developed an evaluation framework for large language models, benchmarking performance across business domains and client-specific requirements.
MARTECH TRIBE
SEPT 2022 – SEPT 2023Implemented proprietary algorithms for vendor matching using SVMs, BERT, and Deep Neural Networks to optimise operational efficiency.
TRAINING PROTOCOLS
VRIJE UNIVERSITEIT AMSTERDAM
LOADED MODULES — CLICK TO EXPAND
PROGRAM PARAMETERS
ML / AI
LANGUAGES
FRAMEWORKS & LIBRARIES
TOOLS & PLATFORMS
EXECUTED PROGRAMS
Deep-Q Learning for Schnapsen
Developed a custom DQN agent using PyTorch for discrete action spaces. Built a reinforcement learning environment with self-play, experience replay, and Huber loss optimisation.
Spectral LoRA
Developing a novel fine-tuning method that adapts Low-Rank Adaptation (LoRA) across transformer layers according to each layer's intrinsic dimensionality. By applying Singular Value Decomposition to measure stable rank, the method allocates variable LoRA ranks per layer — achieving more efficient fine-tuning while deepening understanding of how transformer models represent information. Sits at the intersection of applied ML and theoretical insight, reflecting a move from engineering systems to research-informed development.
OFF-GRID ACTIVITIES
OPEN CHANNEL
Ready to collaborate? Transmit your message through the grid. All communications are logged by the MCP.