PROJECT 15•AI ENGINEERING•AI / ML

AI Code Review Assistant & MLOps Pipeline

AI-powered GitHub Action for automated pull request code reviews using a fine-tuned LLM with LoRA

PythonLoRAHugging FaceDockerGitHub ActionsCI/CDMLOpsLLM Engineering

Integration

GitHub Action

Automated PR review trigger

Fine-Tuning

LoRA PEFT

Parameter-efficient style adaptation

Inference

Dockerized

Lightweight containerized model serving

Feedback

Inline Comments

Context-aware code diff annotations

// INTERACTIVE SYSTEM TELEMETRY & DIAGNOSTIC LAB

RUNTIME ENGINE & LATENCY BENCHMARK COMPARATOR

Empirical benchmark comparing INT8 Post-Training Quantized ONNX against vanilla TorchScript C++ tracing.

INFERENCE BATCH SIZE:

P95 Latency

24.8ms

Deterministic SLA

Throughput

40.3 FPS

Video streaming limit

RAM Footprint

14.2 MB

Model weight & graph

CPU Usage

38%

8-Core Edge node

Target: Sub-30ms budget on edge hardware✓ 3.1x Faster Than TorchScript

01 // SYSTEM OVERVIEW

Code reviews are essential for software quality but consume significant senior engineering bandwidth. This project implements an AI-powered GitHub Action that automatically analyzes pull request diffs, flags security vulnerabilities and anti-patterns, and posts actionable inline comments using an LLM fine-tuned with Low-Rank Adaptation (LoRA).

02 // THE PROBLEM & ENGINEERING SIGNIFICANCE

The Core Challenge

Manual PR reviews often create engineering bottlenecks, delaying deployments while senior developers catch routine syntax, security, or style issues.

Why This Matters

Automating initial code review passes frees senior engineers to focus on high-level architecture while maintaining rigorous codebase standards.

Key Constraints:

•Keeping review comments concise, actionable, and free from repetitive hallucinated criticisms.
•Parsing unified git diffs and mapping comments to precise line numbers.
•Minimizing inference latency so PR checks complete within standard CI/CD time budgets.

03 // DATA PIPELINE & PREPROCESSING

Input Format: Git diff patches, PR metadata, and repository style guidesSample Volume: Curated code review dataset with real-world senior engineer PR feedback

Transformation Steps:

Diff parsing and tokenization filtering out generated lockfiles and assets
Context window compaction retaining hunk headers and surrounding scope lines
Prompt template structuring with zero-shot formatting constraints

Cleaning Strategy: Filtering subjective or conversational comments to retain only concrete bug fixes and style corrections.

04 // SYSTEM ARCHITECTURE & DATA FLOW

GitHub PR Webhook -> GitHub Action Runner -> Dockerized Inference Container -> LoRA Model -> Actionable Markdown Review -> GitHub PR Comments API.

STEP 01GitHub Actions · YAML

GitHub Action Workflow

Triggers on pull_request events and extracts the modified git diff hunks.

STEP 02Python

Diff Preprocessor

Cleans diffs, removes lockfiles, and constructs structured evaluation prompts.

STEP 03PyTorch · PEFT · Hugging Face

LoRA Fine-Tuned Model

Language model adapted on high-quality code review datasets.

STEP 04GitHub REST API

Review Commenter

Posts inline markdown suggestions and summary reviews via GitHub REST API.

05 // MODEL ENGINEERING & HYPERPARAMETERS

Base Architecture: LoRA (Low-Rank Adaptation) on Instruction-Tuned LLM

Fine-tuned with rank r=16, alpha=32, target modules [q_proj, v_proj] using Hugging Face PEFT.

Hyperparameters & Training Dynamics:

• LoRA Rank: 16
• LoRA Alpha: 32
• Learning Rate: 2e-4
• Epochs: 3

Loss Function: Causal Language Modeling Cross-Entropy Loss

Trade-off Rationale: Fine-tuned lightweight adapters (sub-50MB) rather than full parameter weights, drastically lowering hosting and CI cost.

06 // FAILURE ANALYSIS & ZERO-TRUST SAFEGUARDS

OBSERVED FAILURE MODES UNDER STRESS

• Massive pull requests modifying 1,000+ lines exceeding LLM context windows.
• Exotic language syntaxes not present in fine-tuning corpus.

Mitigation & Fallback: Hunk chunking with priority ranking (focusing on business logic files and excluding generated files).

07 // PRODUCTION DEPLOYMENT SPECS

Serving Framework

Containerized Inference Runner

Containerization

Docker image deployed as a reusable composite GitHub Action

P95 SLA

12.4s per standard 200-line diff review

Throughput

On-demand execution via GitHub Actions runners

08 // ARCHITECTURAL DECISIONS & TRADE-OFFS

Used LoRA parameter-efficient fine-tuning rather than pure prompt engineering.

Why: LoRA embeds strict domain-specific review tone and brevity directly into weights without wasting precious prompt context tokens.

Alternative Discarded: Long few-shot prompt injection.

Packaged as a native GitHub Action composite runner.

Why: Allows developers to add the automated reviewer to any repository with a simple 5-line YAML workflow configuration.

Alternative Discarded: Self-hosted webhook server.

09 // PLANNED IMPROVEMENTS & NEXT REVISIONS

→Incorporate repository-wide static analysis (AST parsers) as an additional context layer for the LLM.
→Add automated one-click 'Apply Suggestion' GitHub patch generation.

BACK TO ALL PROJECTS

INSPECT ON GITHUB