My Projects
Selected work in AI research, systems, and creative applications.
ELODIN: Naming Concepts in Embedding Spaces
During my time as a Computer Vision Research Scientist at Modelme, I worked on generating photorealistic virtual fashion photoshoots using diffusion models. I developed a novel method using custom CLIP-based embeddings to achieve better consistency in pose and angle across generated images of the same virtual model. I was involved from the initial research and ideation through to production deployment of the image generation pipeline. This work resulted in the research paper "ELODIN: Naming Concepts in Embedding Spaces" (arXiv:2303.04001).
- Focus: Research and production of consistent generative imagery using diffusion models
- Technologies: PyTorch, Stable Diffusion, CLIP, Docker, OpenAPI
- Link: arXiv:2303.04001
Hytale Code Explainer
A RAG system built over the decompiled Hytale server Java codebase to help with documentation and modding. The project decompiles the game’s JAR file, processes and embeds the code, and allows users to ask natural language questions about the codebase. It supports multi-turn conversations and returns answers with references to specific files and lines. The system was deployed and actively used via a public Discord bot.
- Focus: Production RAG over a large decompiled Java codebase
- Technologies: Python, Qdrant, LiteLLM, VineFlower, Docker, Clean Architecture
- Link: GitHub
Prompt Warrior
An experimental roguelike game built in Godot with deep integration of large language models. Player actions, world state and game rules are interpreted and resolved in real time using diffusion-based text models. Every stage is defined completely in natural language, resulting in a fluid and creative experience for players and builders.
- Focus: Exploring simulation capabilities of LLMs in custom, non-common-sense scenarios such as a game world
- Technologies: Godot (GDScript), LLM APIs, Diffusion Text Models (Mercury 2)
- Link: Live Demo
Latent Translation
A technical project exploring the translation of latent representations between different generative models. An MLP was trained to map latents from EffNet (Stable Cascade) into CLIP’s latent space, enabling semantic comparisons without processing images through the pixel domain.
- Focus: Efficient CLIP comparison in latent space
- Technologies: PyTorch, EffNet (Stable Cascade), CLIP, Jupyter Notebooks
- Link: GitHub
AI Rooms
A text-based escape room game where players interact with the environment using freeform natural language. Each object in the room functioned as its own autonomous agent. The system used structured generation to maintain consistency across agent behaviors and game state while allowing creative player input.
- Focus: Multi-agent LLM systems in interactive environments
- Technologies: Python, React, FastAPI, OpenAI API, Structured Generation
- Link: GitHub