20 years of engineering experience, now applied to bridging traditional game engines (C++ / Unreal) with neural world models (PyTorch).

Research focus: real-time world models and ethical data synthesis.

Interested in LLM-assisted tooling and prompt engineering as a way to revisit and extend classic game experiences.

Active professional work is primarily in private organizational repositories.

This profile will contain selected public artifacts and demos.

Note: The underlying architecture is proprietary. Below are outputs demonstrating the system's ability to synthesize worlds and infer control states. This was entirely a team effort, of some of the most talented people I ever had the pleasure to work with. Everything was trained on a single local 4090 for v1 and a single local 5090 for v2 and is capable of inference locally at 20fps on a single 4090 (imagine if we had the resources to go bigger.)

• Goal: Impressed by sudden emergence of world models, my team coalesced around the idea of creating our own proprietary version of the technology.
• Constraint: Optimized to train entirely on a single consumer GPU (RTX 4090).
• Pipeline Optimization: To solve training bottlenecks, we decoupled the data pipeline. We pre-encoded inputs and frames into tensor chunks, effectively splitting the process to maximize throughput.
• Validation: Initially benchmarked against Minecraft data to verify our novel architecture (distinct from SOTA), then validated on a barebones UE5 environment.

• Goal: To establish a fully ethical, high-fidelity data source. This version utilized an expanded architecture to handle increased environment complexity.
• Constraint: Optimized to train on a single RTX 5090 while maintaining inference capability on a 4090.
• Data Harvesting: To solve data volume bottlenecks, we deployed bots with multiple viewports, traversing the game world to capture 4x more data per instance.
• Validation: The architecture proved highly capable, though the results highlighted an exponential need for data scaling to eliminate artifacts.

To ensure 100% data ownership and avoid excessive web-scraping, we authored a custom "Data Laboratory" in Unreal Engine 5.

• Process: Automated agents navigate a custom-built 3D world to capture perfectly synchronized Frame + Input pairs.
• Objective: To train the Predictor on this ground-truth data, enabling future "zero-shot" auto-labeling of external video sources.

Testing whether modern LLMs can accelerate game development workflows. Each prototype below was built from scratch in around 10 minutes using various available LLMs to port game logic to standalone HTML/JS implementations.

Goal: Validate that LLMs enable developers to test wildly different game concepts without investing days per prototype.

Moody first-person raycaster with shooting mechanics.

▶️ Play Live Demo

First-person raycaster with multiple characters and RPG systems.

▶️ Play Live Demo

OutRun-inspired racing with vintage-style sprite scaling.

▶️ Play Live Demo

Development time: ~10 minutes each | Stack: Pure HTML5/JS, no frameworks | Workflow: Iterative prompting Claude, Gemini, and ChatGPT

• Project: ComfyUI-itsB34ST-Nodes
• Focus: Custom node systems for advanced generative workflows, including stylization and animation control.

• Project: Modern-Neuro-Stylize
• Focus: Modernized neural style transfer experiments with custom architecture modifications.

• Languages: Python, JavaScript/TypeScript, C++, HLSL
• Machine Learning: PyTorch, Custom Architecture Design, Synthetic Data Pipelines
• Engines: Unreal Engine 5, Custom Neural Renderers
• **Community:**Senior Moderator at Banodoco (ML Community)