StealthHumanizer

🥷 Free, open-source AI text humanizer — corpus-trained on 10,000 Q1 academic papers. 13 AI providers, 4 rewrite levels, multi-pass ninja mode. No login, no limits, 100% client-side.

Live app: https://stealthhumanizer.vercel.app/ · Docs: https://rudra496.github.io/StealthHumanizer/

Table of Contents

• Features
• How It Beats AI Detectors
• Architecture
• Installation
• Quickstart
• Configuration
• Usage Examples
• Testing and Local Development
• Benchmarks and Performance
• Roadmap
• Contributing
• Security, Support, and License

Features

• Corpus-trained humanization engine built from 10,000 Q1 open-access academic papers spanning 11 domains (2018–2025).
• Dynamic detection thresholds calibrated against real human writing patterns, not guesswork — sentence length, burstiness, vocabulary diversity, and transition frequency.
• Expanded AI phrase database with 150+ collocation replacements for natural output.
• Domain-aware style matching across 11 academic disciplines.
• 13 AI provider support with configurable API keys (free and paid).
• 4 rewrite levels including multi-pass ninja mode for maximum transformation.
• 13 preset writing tones and granular style controls.
• Integrated AI detection with corpus-calibrated heuristic scoring.
• Readability analysis (Flesch-Kincaid, Gunning Fog).
• PDF and DOCX file upload support.
• Grammar check integration.
• Multi-language humanization support.
• Side-by-side workflow for source, output, and quality feedback.
• Browser-first key handling — all API keys stay on your device.
• Dark/light theme toggle.

How It Beats AI Detectors

StealthHumanizer uses a multi-layer approach grounded in real data, not heuristics:

1. LLM rewrite — your chosen provider transforms the text using a corpus-aware prompt injected with statistical targets from 10,000 real Q1 papers.
2. Corpus-aware post-processing — an expanded collocation engine replaces 150+ known AI-signature phrases with natural alternatives.
3. Detection calibration — the built-in detector scores output against dynamic thresholds derived from real human writing (sentence length mean 20.5, burstiness 0.426, vocabulary diversity 69.4%, passive voice 18.1%).

The result is text that doesn't just avoid AI patterns — it matches human writing patterns measured from actual published research.

Architecture

High-level architecture:

1. UI layer (components/, app/page.tsx) — text entry, settings, and result rendering.
2. API routes (app/api/) — provider orchestration and rewrite workflows.
3. Style model layer (public/corpus-style-model.json) — corpus statistics and calibrated thresholds loaded client-side from 10,000 Q1 papers.
4. Core logic (lib/) — prompt construction (with corpus-aware injection), provider abstraction, detector scoring, and storage helpers.
5. Research and evaluation scripts (scripts/, data/) — benchmark, training, and corpus ingestion pipelines.
6. Documentation (docs/) — user and contributor guides published via GitHub Pages.

For deeper technical details, see ARCHITECTURE.md and STYLE_ENGINE.md.

Installation

Prerequisites

• Node.js 20+
• npm 10+

Setup

git clone https://github.com/rudra496/StealthHumanizer.git
cd StealthHumanizer
npm ci

Quickstart

Run the application locally:

npm run dev

Then open http://localhost:3000, add a provider API key in settings, and run a rewrite.

Configuration

StealthHumanizer is configured primarily through UI controls and local browser storage.

• Provider keys: configured in app settings and stored locally.
• Rewrite strategy: choose level, style, tone, and target score.
• Research pipeline scripts: use JSON configs under data/papers/*.config.example.json and data/models/*.config.example.json.

See docs/configuration.md for full details.

Usage Examples

Application usage

1. Paste AI-generated text.
2. Select rewrite level, style, and tone.
3. Run humanization.
4. Review detector/readability scores and iterate.

One-command corpus + training bootstrap

npm run pipeline:q1-ready

Faster reruns (skip reinstall):

npm run pipeline:q1-ready:skip-install

The wrapper (scripts/papers/complete-ready-pipeline.mjs) runs:

1. npm ci (unless --skip-install)
2. node scripts/papers/batch-download-and-train.mjs with Q1 OA configs
3. node scripts/model/evaluate-framework.mjs --manifest data/models/current/run.manifest.json

Scripted benchmark/training smoke flow

npm run papers:benchmark -- --config data/papers/benchmark.smoke.config.json --run-id local-smoke
npm run model:train -- --config data/models/train.smoke.config.json --run-id local-smoke
npm run model:eval -- --manifest data/models/current/run.manifest.json

Testing and Local Development

npm run lint
npm run test:integration
npm run build

See CONTRIBUTING.md for workflow standards.

Benchmarks and Performance

• docs/BENCHMARK_METHODOLOGY.md
• docs/HUMAN_AI_SIGNAL_LAYER.md
• docs/MODEL_TRAINING_FRAMEWORK.md

Roadmap

See ROADMAP.md for release milestones and planned improvements.

Contributing

Contributions are welcome. Please review CONTRIBUTING.md before opening a pull request.

👨‍💻 Author

Rudra Sarker — 3rd-year IPE student at SUST, Bangladesh. Building open-source tools for accessibility, education, and developer productivity.

Security, Support, and License

• Security policy: SECURITY.md
• Support channels: SUPPORT.md
• Code of Conduct: CODE_OF_CONDUCT.md
• License: MIT