OG
Om Ghante
•FOCUSCurrently working on ScheduleSomething and LayerdEvents algorithmsLearn More
•FOCUSCurrently working on ScheduleSomething and LayerdEvents algorithmsLearn More
•FOCUSCurrently working on ScheduleSomething and LayerdEvents algorithmsLearn More
•FOCUSCurrently working on ScheduleSomething and LayerdEvents algorithmsLearn More
Back to Projects
AI & MLOpen Source

Patch.ai

Autonomous software reliability & self-healing AI engine performing PyTorch GNN-driven Root Cause Analysis on runtime crash stack traces to generate safe patch proposals.

⚡ Impact Outcome: Constructs incident-code graphs to diagnose runtime stack trace failures and propose automated self-healing patches.
View GitHub Repository Live Demo Read GitHub Docs

Tech Stack

Python 3.10+PyTorchNetworkXGraph Neural NetworksClick CLIWatchdogHugging Face Hub

Architecture Concepts

PatchGNN Neural NetworkIncident-Code GraphAutonomous RCALive Telemetry Watcher

Live GitHub Repository README

Live Synced

Patch.ai — Autonomous Reliability & Self-Healing AI Engine

Patch.ai acts as a production immune system for software applications. It ingests application runtime crashes, error stack traces, and telemetry, builds a heterogeneous Incident-Code Graph, performs PyTorch GNN-driven Root Cause Analysis (RCA), and generates safe, risk-scored patch proposals.

Ecosystem Synergy: Reecall.ai + Patch.ai

   Developer ───► Reecall.ai (Development Memory / Architectural Context)
                       │
                       ▼
                 System Build
                       │
                       ▼
    Production ──► Patch.ai (Runtime Memory / Autonomous Self-Healing)
  • Reecall.ai: Preserves code context, architecture decisions, AST structure, and developer intent.
  • Patch.ai: Monitors runtime failures, performs Graph Neural Network RCA, predicts root causes, and proposes verified patches.

Directory & Package Architecture

Patch.ai uses a modular Python package structure mirroring 
reecall.ai
:
patch.ai/
├── cli/                # Command line interface (`patch-ai`)
├── configs/            # Configuration defaults
├── dataset/            # Incident dataset generator & PyTorch dataloaders
├── features/           # Failure heuristics, feature mapping, & crash clustering
├── graph/              # Heterogeneous Incident-Code Graph builder & store
├── inference/          # Root Cause Search, retriever, & risk ranker
├── memory/             # Patch prompt context builder & risk scorer
├── model/              # PyTorch PatchGNN neural network & vector embeddings
├── parser/             # Stack trace scanner, AST symbol parser, import analyzer
├── patch_ai/           # Package initialization
├── tests/              # Pytest suite
├── watcher/            # Live telemetry log stream observer
├── train.py            # End-to-end GNN training pipeline
├── train_dataset.py    # Synthetic crash incident dataset generator
├── infer.py            # CLI inference entrypoint for stack trace RCA
├── eval_model.py       # Model evaluation & RCA recall benchmark
└── upload_hf.py        # Hugging Face Hub model uploader

Installation & Setup

bash
git clone https://github.com/om-ghante/patch.ai.git
cd patch.ai

# Install in editable mode
pip install -e .

Quick Start

1. Generate Synthetic Incident Dataset

bash
python train_dataset.py --samples 500 -o data/dataset.json

2. Train PatchGNN Model

bash
python train.py /path/to/target/repo --epochs 50 --output data/checkpoints

3. Diagnose Crash Trace & Suggest Patch

bash
python infer.py /path/to/target/repo "ZeroDivisionError in calculate_metrics" --top-k 3

CLI Usage

bash
# Scan repository and build incident graph
patch-ai scan /path/to/repo

# Run RCA on a stack trace file
patch-ai diagnose /path/to/repo /path/to/trace.log

# Generate patch proposal
patch-ai patch /path/to/repo /path/to/trace.log

# Monitor live log directory for runtime exceptions
patch-ai watch /path/to/logs

License

MIT License © 2026 Om Ghante.