01
Lizard / Halstead Analyzer
Baseline
Establishes cyclomatic complexity and structural signatures of the original code.
JAVAREFACTORING STUDIO
Runs fully local with no cloud dependency or API keys. Targets consumer GPUs (as low as 4GB VRAM) with automatic CPU fallback.
01 / Setup
INSTALLATION
No build required. Everything (app code + models) is bundled inside the pre-built image and pulled from GitHub Container Registry.
Docker-readyLocal-firstPre-built modelsNo telemetry
What happens next?
01 / Copy & Run
Run the compose command in your terminal to automatically pull the pre-quantized models and configure local VRAM serialization.
02 / Access Studio
Navigate to http://localhost:3000 to load the desktop refactoring studio workspace.
03 / Refactor
Upload your Java source files. The multi-agent pipeline immediately kicks off analysis, displaying live AST changes in the workspace.
02 / Pipeline
PIPELINE PHASES
02
Planner Agent (Qwen2.5-Coder 3B)
Strategy
Classifies refactoring intent and synthesizes an AST modification plan.
03
Generator Agent (Qwen2.5-Coder 3B)
Execution
Applies mutations sequentially or single-shot, generating multiple scored variants.
04
javalang AST Validator
Validation
Runs syntax, complexity-gate, and boundary checks to catch unintended side effects.
05
Judge Agent (Llama-3.2 3B)
Adjudication
Audits the change for semantic correctness and overrides hallucinated diffs.
06
Insights & Save
Finalization
Cleans up, persists results, and extracts developer-facing insights.
01
Lizard / Halstead Analyzer
Baseline
Establishes cyclomatic complexity and structural signatures of the original code.
02
Planner Agent (Qwen2.5-Coder 3B)
Strategy
Classifies refactoring intent and synthesizes an AST modification plan.
03
Generator Agent (Qwen2.5-Coder 3B)
Execution
Applies mutations sequentially or single-shot, generating multiple scored variants.
04
javalang AST Validator
Validation
Runs syntax, complexity-gate, and boundary checks to catch unintended side effects.
05
Judge Agent (Llama-3.2 3B)
Adjudication
Audits the change for semantic correctness and overrides hallucinated diffs.
06
Insights & Save
Finalization
Cleans up, persists results, and extracts developer-facing insights.
03 / Features
CORE FEATURES
Built for fully local, privacy-first, automated refactoring.
Architecture
Multi-agent orchestration
Three small language models (planner, generator, judge) collaborate to produce high-quality refactoring.
Pipeline
6-phase refactoring pipeline
Plan, generate, verify, mutate, deduplicate, and rank in one automated workflow.
Deployment
GPU + CPU support
A single pre-built image auto-detects GPU at runtime; machines without a GPU run CPU-only automatically.
Communication
Real-time WebSocket
Live streaming of analysis, code diffs, and progress updates to the UI.
Privacy
No cloud dependency
Everything runs fully local on commodity hardware; no API keys, no data leaving the machine.
Integration
REST API
A full HTTP API for headless integration and automation.
04 / Studio Demo
REFACTORING IN ACTION
See how the SLM orchestration translates complex, nested conditionals into structured, clean, and optimized Java patterns.
Input.java
# 17 lines
1234567891011121314151617
public class OrderTracker {
public String getStatusMessage(int s) {
String msg = "";
if (s == 0) {
msg = "Order Placed";
} else if (s == 1) {
msg = "Processing";
} else if (s == 2) {
msg = "Shipped";
} else if (s == 3) {
msg = "Delivered";
} else {
msg = "Unknown";
}
return msg;
}
}
Ready
12345678910111213
public class OrderTracker {
private static final String[] messages = {
"Order Placed",
"Processing",
"Shipped",
"Delivered"
};
public String getStatusMessage(int s) {
return messages[s];
}
}
05 / Principles
ARCHITECTURE
FULLY LOCAL
Zero cloud dependency. Runs fully local using quantized GGUF models.
DETERMINISTIC OUTPUTS
Grammar-constrained generation (GBNF) forces structural JSON alignment, backed by deep syntactic and semantic validation.
FEEDBACK LOOPS
Graceful failure logic. Failed phases accumulate structured feedback to guide the next retry instead of failing silently.
SERIALIZED EXECUTION
Single concurrency execution. Heavy model operations are serialized to prevent VRAM overflow on consumer hardware.
06 / Team
Meet the builders
This is the team that built the Multi Agent Orchestration system behind Horizon, driving automated AST checks, telemetry flows, and refactoring pipelines.
Project Lead
Orchestrates overall direction, multi-agent workflows, and coordination across the refactoring pipeline to align local agent orchestration with system goals.
View portfolio›Want to refactor your java code fully offline and no ai cloud dependency?
Deploy Horizon locally inside your secure intranet environment using pre-built GGUF models.
Privacy-first, local multi-agent LLM orchestration for an automated Java AST refactoring pipeline.
Architecture
- AST Verification
- GBNF Grammars
- Quantized GGUF
- Feedback Loops
Builders
- Christian BalinadoProject Lead
- Joshua LopezFull Stack / Lead Developer
- Jericho VardeFrontend Developer
- Andrew DejitoQuality Assurance Manager
© 2026 Horizon AI.