The AI Architectural Linter

A per‑repository micro‑model that understands your project’s architecture, patterns, and intent - and protects it from drift, rot, and inconsistency.

The Problem: AI Drift

AI accelerates development, but it also accelerates architectural drift. Inconsistent patterns, duplicated logic, and context‑free code generation are becoming the new technical debt.

Context Limits

Even the best AI models only see a slice of the codebase. They lack persistent, repo‑specific memory and cannot enforce long‑term structure.

Read more

Inconsistent Patterns

Different agents and developers introduce different approaches. Naming conventions drift, and modules evolve independently instead of cohesively.

Read more

Review Overload

Senior engineers spend too much time policing patterns and explaining architectural decisions that should be self‑evident.

Read more

The Solution: A Per‑Repo Micro‑Model

ArchGuardian introduces a project‑specific AI model that lives inside your repository and evolves with every change.

1

Micro‑Model Per Repository

Small enough to train locally (1–3B), but large enough to encode patterns, naming, and architecture.

Learn More

2

LoRA Updates on PR Merge

A GitHub Action fine‑tunes the model using the diff, context, PR description, and review comments.

Learn More

3

Scheduled Rebuilds

Nightly or weekly jobs rebuild the model from scratch to prevent drift and ensure full context.

Learn More

Why This Works

  • Persistent Context: The model remembers architectural decisions.
  • Privacy First: Models stay in your repo; no cross‑tenant training.
  • Agent Agnostic: Works with Copilot, Claude, Cursor, or any MCP tool.
  • Economically Viable: Client‑side training means no massive hosting costs.

System Architecture

How ArchGuardian integrates into your existing workflow.

1. Base Model

Standard 1-3B parameter model suitable for LoRA.

View Details →

2. Fine-Tuning

GitHub Actions run LoRA updates on every PR merge.

View Details →

3. Inference

Agents query the model via MCP for architectural guidance.

View Details →

Example Interactions

See how ArchGuardian acts as an intelligent architectural assistant.

Pattern Detection

"Where is OAuth implemented in this repo?"

> Library: Spring Security
> Pattern: Custom middleware
> Entry point: AuthConfig.java
> Used in: gateway, user-service

Impact Analysis

"What breaks if I change UserDTO?"

> 3 Service classes affected
> 2 API endpoints need updates
> 5 Test suites will fail

Architectural Linting

"Is this PR consistent with our event handler pattern?"

> WARNING: Event naming mismatch.
> Expected: [Domain]Event
> Found: [Domain]Message

Canonical Guidance

"How do I add a new event handler?"

> Use the @EventHandler annotation.
> Extend BaseEvent.
> Register in EventRegistry.java.

FAQ

Why not just use embeddings or RAG?

Embeddings provide retrieval, not understanding. They cannot learn patterns, enforce architecture, or evolve with the codebase in the same way a fine-tuned model can.

Is this secure?

Yes. Models are stored inside your repository. Training happens locally or on self‑hosted runners. There is no cross‑tenant training or centralised model hosting.

Does this replace Copilot?

No. ArchGuardian is a context provider. It enhances Copilot, Claude, and other agents by giving them the deep, project‑specific architectural intelligence they currently lack.