Skip to content
Formray
Torna a Field Notes
Build Log3 min di lettura

Introducing PRISM

Giuseppe Albrizio/

PRISM is a desktop application for detecting and humanizing AI-generated text. It runs locally, analyzes offline, and gives you control over how your writing sounds.


The problem

AI-generated text has a signature. Uniform sentence lengths. Predictable transitions. Hedging language. Corporate jargon. No personality, no voice, no humanity.

The Real Issue

The problem isn't that people use AI to write. The problem is that AI-generated text is identifiable — and nobody has good tools to understand why or where in their text the signal shows.

Writers need to know if their drafts sound robotic. Editors need to spot AI-assisted submissions. Professionals need to polish AI-drafted content without sounding like a machine.

The existing detection tools are cloud-based black boxes. You paste your text into someone else's server, get a percentage back, and hope they're not training on your content.


The approach

PRISM analyzes text locally using a combination of heuristics and machine learning.

Cloud DetectorsPRISM
Text uploaded to external serversEverything runs on your machine
Black box score, no explanationSentence-level heatmap with reasons
Subscription for basic featuresOpen source, MIT licensed
Single language, usually EnglishEnglish, Italian, Spanish, German

Perplexity, burstiness, entropy, and pattern matching work together to produce an AI probability score. A sentence-level heatmap shows exactly which parts triggered detection and why.


Two modes

1

Detect

Paste text. View the AI probability score and metrics. See a color-coded heatmap highlighting which sentences triggered detection. Click humanize for three style variations — Natural, Creative, Professional.

2

Compose

Write rough text in any language. Select source and target language. Get three polished versions with distinct voices. Copy the one that fits. Repeat and refine.


Architecture

Stack

Built with Tauri 2.0 — a lightweight Rust-based desktop framework. The React frontend handles UI and state management through Zustand, while all analysis runs natively in Rust.

Text analysis combines ONNX-based machine learning inference with heuristic metrics. The Claude API provides humanization when configured. API keys are stored in the OS keychain — never in config files.

The design decision that shaped everything: local first.

Detection runs entirely offline through ONNX Runtime. No internet required. No data leaves your machine. Humanization can use the Claude API for the highest quality, or local ONNX models for complete offline operation.


Why open source

A detection tool that isn't transparent about its methods isn't a detection tool. It's an oracle asking you to trust it.

PRISM is MIT licensed. The detection logic, the heuristic weights, the ML model — all visible, all auditable. If you disagree with how we score text, you can see exactly why and adjust it.

This matters because AI detection is inherently probabilistic. No tool is 100% accurate. The least we can do is show our work.


Where we are

Status: R&D — v0.4.0

PRISM is functional and open source on GitHub. The detection engine is stable. Humanization works through both Claude API and local ONNX models.

Active development continues on detection accuracy and language expansion.


What's next

Better models, more languages, sharper heuristics. The goal is a tool that helps you write with AI without losing your voice.

See through synthetic text. Write through it.

Torna a Field Notes