Agentic Projects

Eclipse

A complete Unreal Engine 5.7 environment where no human painted terrain, wired materials, or placed a single asset.

I designed the architecture and built every tool in the chain. AI agents operated them. The project splits into two pipelines that share a common 12-layer landscape system: generate terrain once, everything else follows.

Asset Pipeline

Ask a Claude agent for 40 Scottish Highland rocks and go do something else. The agent writes FLUX.2 prompts for each asset, generates concept art in batch, feeds them through Trellis2 for 3D generation, then QPaint for AI texturing from wireframe projections. Truck (custom batch tooling) handles the volume, running assets through headless Blender and Nanite processing into Unreal automatically. A C++ PCG system reads the shared landscape layers and spawns them across 254,016 sample points. Single asset turnaround is 20-30 minutes. ~930 props generated this way for the project, 95%+ batch success rate.

Engine Pipeline

Claude agents direct both Houdini and Unreal through MCP. On the terrain side, a Houdini MCP agent generates the heightfield, erosion, noise, and all 12 terrain layer masks from slope and curvature analysis. A Python script programmatically builds the full 12-layer triplanar material graph with no editor node wiring. On the engine side, Claude works through Unreal MCP for blueprints, C++ code, PCG systems, material setup, and scene assembly. The 12 shared layers tie it all together: Houdini masks, the UE5 material, and the PCG foliage spawner all read the same data.

Unreal Engine 5.7 · C++ · Python · Houdini MCP · Unreal MCP · Claude · FLUX.2 · Trellis2 · Blender

Blukraft

AI digital twin and multi-machine agent, reachable via Discord DM.

A persistent Claude agent running 24/7 on a Mac Studio that acts as a technical co-developer across multiple game projects. Reachable through Discord DM, it coordinates work across three networked machines over Tailscale, spawning sub-agents for parallel tasks and SSHing into an Unreal/Houdini dev rig and a GPU workstation for 3D asset generation. It records and transcribes meetings with speaker identification, manages todos, audits codebases, deploys fix teams, generates images, and reads and writes code across all three machines. It also wrote this description.

Claude Code · Tailscale · Discord · MLX Whisper · Flux · Unreal MCP · Houdini MCP · Trellis2

Battle Aces Lore Agent

Multi-model writing agent that shipped production game copy.

Built a local lore master agent on an M2 Ultra Mac Studio using LangChain and custom Python. A DeepSeek 70B handled creative generation while a Mixtral 8x7B served as the logical connector, keeping output consistent with established canon. The agent used RAG over the game's universe knowledge base and could proofread the live Unity project directly. It wrote final unit descriptions, VO lines, and voice direction notes that shipped in Battle Aces. No cloud, no API costs, just two local models and a well-structured pipeline.

Python · LangChain · DeepSeek 70B · Mixtral 8x7B · RAG · M2 Ultra (local)

The Last Hundred (Play Prototype)

An exploration into a fully agentic game development pipeline resulting in a mobile hero collector RPG.

Built in just 20-30 hours over the holidays working alongside an AI agent, this technical demo proves the viability of AI-accelerated end-to-end production. Most development time was spent scaffolding core systems rather than grinding content. Featuring 30+ minutes of gameplay spanning quests, combat, and dialog, it demonstrates how shifting focus from asset creation to systemic iteration allows a solo developer to ship a vertical slice of a complex genre. Best played on an iOS mobile device.

Claude Code · Godot 4 · Godot MCP · Gastown · Nano Banana Pro (Python)