Diagnostics
Produces a chapter-level critique that shows where the argument breaks down, which sections are overloaded, what feels redundant or underdeveloped, and what should be restructured first.
All work
Scientific editing
AI Book Editor
A right-rail AI review system for scientific and medical manuscripts, with diagnostics, structure analysis, fact-check triage, spellcheck, and targeted block-level edits.
- My role
- Product Owner, AI Architect & Full-stack Engineer
- System
- Next.js, React, TypeScript, CodeMirror, OpenAI, Gemini, Anthropic, LanguageTool
01
Editorial intelligence
Built from scratch around the manuscript editor, right-rail AI review system, task-specific prompts, provider orchestration, deterministic validation, and deployment.
The product is designed for long scientific and medical books where the main problem is not generating more text. The editor needs to know where a chapter loses the reader, which claims are risky, what structure is broken, and which changes are worth making.
ORESTGPT turns a manuscript into separate review layers: diagnostics, structure, fact-checking, spellcheck, clarity, formatting, emphasis, visual suggestions, and local block-level edits.
02
Right-rail AI tools
Each tool produces a different kind of editorial output, so the editor can see what is structurally wrong, what is factually risky, and what is simply hard to read.
Structure analysis
Produces a usable map of the chapter: where section boundaries really are, where subheadings are missing, where transitions fail, and which parts should be split, merged, or reordered.
Fact-check triage
Produces a verification queue for the claims that actually look dangerous or weak: outdated medical classifications, suspicious numbers, unsupported treatment statements, and other assertions that need human checking.
Spellcheck
Produces deterministic spelling and grammar corrections with suggestions and dictionary controls. This runs through LanguageTool instead of an LLM because it is more precise, more consistent, and much cheaper for that job.
Clarity and interest
Produces targeted rewrites for passages that are dense, repetitive, abstract, or flat, helping the author explain difficult ideas more clearly without flattening the original meaning.
Formatting and emphasis
Produces concrete presentation fixes: where prose should become a list, table, callout, subsection, highlighted takeaway, or visual explanation so the chapter becomes easier to scan and understand.
03
Architecture choices
Step-aware prompts
Diagnostics, structure, fact-checking, clarity, formatting, emphasis, and local edits use separate prompts and output contracts, so each AI pass stays inside its editorial responsibility.
Tool selection
LLMs handle reasoning about structure, evidence, clarity, and reader comprehension, while LanguageTool handles grammar and spelling where deterministic tooling is the better fit.
04
Local edits
The editor also supports block-level AI actions such as rewrite, shorten, expand, convert to list, create table, add callout, or generate a visual prompt.
These actions are anchored to selected manuscript blocks and paragraph ranges, so AI output stays bounded and reviewable instead of drifting across the whole chapter.
- Show every recommendation in context before applying it.
- Keep fact-checking separate from style and structure feedback.
- Validate patch application against the current manuscript state.
- Preserve decision history so accepted and rejected changes remain traceable.
05
What the editor gets
Instead of receiving one opaque rewrite, the editor gets distinct outputs for distinct problems: a structural critique, a chapter map, a list of risky claims to verify, deterministic language corrections, and targeted rewrites for weak passages.
The practical result is a chapter that is easier to approve: clearer in structure, safer in factual claims, cleaner in language, and more readable without losing the author's voice.