← Back to Selected Work
SPECIALIZED AI AGENT · GAME DEV & LLM FINE-TUNING

Vivid-Architect (vivid_ue5)

An autonomous Unreal Engine 5 C++ developer agent powered by fine-tuned Qwen2.5-Coder-1.5B with an indexed 57MB Epic Games engine corpus, AST heuristic validation, and a live Poe bot bridge running on local NVIDIA RTX hardware.

Architecture: LoRA Rank-16 SFT + RAG
Compute: Local RTX 3070 Ti (8GB)
Deployment: Poe.com & OpenAI API
Corpus: 57MB UE5 Source Headers
Poe Agent: Vivid86-DevAgent (Active)
Chat with Vivid-Architect on Poe ↗
Vivid-Architect UE5 Developer Agent Overview

The Engineering Challenge

General-purpose frontier LLMs struggle with production Unreal Engine 5 C++. Even models trained on vast quantities of public GitHub code frequently commit severe domain errors: hallucinating deprecated UE4 macros, failing strict Epic Hungarian naming conventions (e.g. failing to prepend A for Actors, U for UObjects, F for structs, E for enums), misplacing the mandatory #include "<Class>.generated.h" header before other includes, and failing to declare module dependencies inside *.Build.cs files.

Furthermore, cloud AI APIs create latency, recurring per-token costs, and privacy concerns when working on proprietary game assets. Vivid-Architect was engineered to solve this: a high-speed, local autonomous agent that understands the strict rules of Unreal Engine 5 and runs on consumer desktop hardware (NVIDIA GeForce RTX 3070 Ti, 8GB VRAM).

Core Architectural Pillars

1. Rank-16 LoRA Fine-Tuning

Fine-tuned Qwen2.5-Coder-1.5B-Instruct over all attention (q, k, v, o) and MLP (gate, up, down) projection matrices using custom-curated UE5 C++ modules covering procedural dungeon generation, spline deformation, Enhanced Input, and GamePlay subsystems.

2. 57MB Engine Corpus RAG

Indexes 1,619 official Unreal Engine source headers and 2,805 classes in ~1 second. Provides real-time signature and type lookups so generated code matches exact engine headers.

3. Autonomous Tooling & Build.cs

Automatically generates paired .h and .cpp files with accurate macro hierarchies and idempotently injects module dependencies (e.g. ProceduralMeshComponent) into the project's .Build.cs.

4. Heuristic AST Validator

Pre-flight static analyzer verifies balanced braces, macro declarations (GENERATED_BODY()), include ordering, and Hungarian notation before files are written to disk.

Fine-Tuning & Quantized Training Pipeline

Training was orchestrated locally with PyTorch and Hugging Face PEFT on a single RTX 3070 Ti (8GB VRAM). The training pipeline utilized 4-bit NormalFloat (NF4) quantization, paged_adamw_8bit optimizer, and gradient checkpointing to fit the Rank-16 adapter within 7.2GB peak VRAM.

# Training Configuration excerpt (train_advanced_lora.py) from peft import LoraConfig, get_peft_model, TaskType lora_config = LoraConfig( r=16, lora_alpha=32, target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], lora_dropout=0.05, bias="none", task_type=TaskType.CAUSAL_LM, )

The resulting fine-tuned adapter (vivid_ue5_lora) exhibits zero syntax drift on Hungarian notation and correctly orders UE5 header includes 100% of the time across test benchmarks.

Poe.com Server Bot Integration

To provide immediate, zero-friction access to the fine-tuned agent from any browser, phone, or tablet, Vivid-Architect is integrated into **Poe.com** via a high-performance bridge:

// Multi-tiered serving topology [Poe.com User] → [Cloudflare Tunnel (HTTPS)] → [FastAPI Poe Bridge (D:\poe-bot\poe_server.py)] → [Local GPU Model Server (http://localhost:8088/v1)] → [Qwen2.5-Coder-1.5B + LoRA on RTX 3070 Ti]

The FastAPI bridge implements fastapi_poe.PoeBot with streaming token support, graceful timeout handling, and automatic status failovers.

Poe Bot Name: Vivid86-DevAgent
Try asking it: "Scaffold a procedural dungeon grid generator actor with InstancedStaticMeshComponent and Enhanced Input" on Poe!

Dual Interface: CLI REPL & OpenAI API

In addition to Poe, Vivid-Architect provides: