AI Algorithm Engineer passionate about building intelligent systems and pushing the boundaries of AI through open-source projects and hands-on experiments.

• 💼 Primary Role: AI Algorithm Engineer working on LLM Fine-Tuning, RAG Systems, Autonomous Agents, and Full-Stack AI Application Development.
• 🧠 Current Research & Focus: Exploring the intersection of World Models, Embodied AI, and Vision-Language-Action (VLA) architectures for robotics.
• 🌱 Philosophy: Building lightweight, reproducible, and impactful proof-of-concepts outside of daily production pipelines.

• AI / ML: PyTorch, Hugging Face, LLM Fine-Tuning (LoRA/QLoRA), RAG, Flow Matching
• Embodied AI & World Models: VLA Architectures, LeRobot, V-M-C World Models, Latent Dynamics
• Engineering: Python, Full-Stack Development (FastAPI, WebSockets, Frontend/Backend), Vector DBs

Embodied AI • VLA • LoRA • LeRobot • Flow Matching

• Overview: Exploration and fine-tuning demo for Hugging Face's SmolVLA (~450M parameter lightweight vision-language-action model based on SmolVLM2-500M and a flow-matching action expert).
• Key Highlights:
  • Fine-tunes lerobot/smolvla_base using LoRA under the Hugging Face LeRobot framework.
  • Trainable and deployable on consumer GPUs for robotic manipulation tasks.
  • Features a series of probe experiments verifying that the LoRA adapter is actually trained, saved, properly loaded, and genuinely alters output behavior during inference.

World Models • V-M-C Architecture • Reinforcement Learning • Generative AI

• Overview: A faithful open-source reproduction of the landmark paper [World Models] by David Ha and Jürgen Schmidhuber (2018).
• Key Highlights:
  • Implements the full V–M–C (Vision – Memory – Controller) architecture.
  • Demonstrates how an autonomous agent learns to compressed representations of the world (Vision/VAE), predicts future dynamic states (Memory/MDN-RNN), and makes decisions inside its own mental simulation (Controller).

Green AI • Dynamic Neural Networks • Sustainability • NAS

• Overview: A proof-of-concept demonstrating that a single fixed AI model shouldn't be the only operational choice.
• Key Highlights:
  • Constructs a dynamic family of neural models instead of relying on a single static architecture.
  • Adapts model execution by switching between variants based on live clean vs. dirty electricity grid supply.
  • Achieves up to ~77.5% reduction in carbon footprint on image datasets without discarding operational utility.

• 📂 GitHub: @UlyssesIE
• 💬 Feel free to reach out for discussions on World Models, Robotics/VLA, or Green AI!