grai

Overview

grai is a local-first HTML5 writing workbench designed to enforce personal writing style and cadence when drafting content with artificial intelligence. Developed as an offline-capable browser utility, the tool extracts a mathematical voice profile from previous writing samples and audits drafts against specific syntactic and rhythm rules. It operates alongside existing large language models without requiring recurring subscriptions, cloud databases, or platform credentials.

Key Features

• Voice DNA extraction — Ingests 5 to 20 writing samples to calculate an author's quantitative fingerprint, including median sentence length and standard deviation.
• Deterministic cadence gate — Audits drafts outside of the language model against six structural rules and enforces a sentence-length coefficient of variation floor of at least 0.32.
• Premise sparring layer — Intercepts and challenges generic or consensual theses prior to drafting to sharpen the core argument.
• Targeted span repair — Flags predictable syntactic patterns such as contrastive negations and flat cadences with visual highlights for manual editing.
• Local-first architecture — Executes entirely inside the browser using local storage or session memory, keeping drafts and voice profiles on the user's machine.
• Flexible model support — Operates via copy-paste workflows across Claude, ChatGPT, and Gemini, with custom build support for direct API keys and local Ollama instances.

Who It's For

grai is built for startup founders, executive ghostwriters, and boutique content agencies publishing commentary and thought leadership on professional networks. It serves operators who rely on generative AI for speed but want to eliminate predictable, formulaic model cadences under their byline. The software is also suited for privacy-focused writers who refuse cloud-based social media schedulers or credential sharing.

Notable Strengths

A primary differentiator is grai's local-first architecture and one-time lifetime license model, which eliminates ongoing subscription fees and prevents user drafts from being stored on external vendor servers. Rather than relying on simple vocabulary blocklists or vague prompt adjectives, the engine enforces deterministic mathematical benchmarks, such as burstiness coefficient thresholds and syntactic pattern checks. Furthermore, because it functions as an external audit layer, it remains agnostic to the underlying generative model being used.