Flyleaf is an innovative web app designed to enhance writing evaluation by distinguishing AI-like text from human prose. This tool aids students, teachers, and writers in refining their drafts without relying solely on numerical scores.
Flyleaf
A first look at a draft before anyone treats a percentage like a verdict
Open FlyleafMost writing tools either score a draft as if the number were proof, or rewrite it as if wording could erase how it was made. I built Flyleaf to sit between those habits. You bring English prose, a PDF, or a Word file. The app marks the share that reads as AI-like, checks public pages for similarity, and can return a plainer draft. The result is an estimate you can discuss. It does not establish that a chatbot wrote the text, and it does not grade the work.
The Problem I Wanted to Solve
A draft can sound smooth for many reasons. A student may have used a model. A careful writer may have edited until every sentence lands the same way. A non-native speaker may write in a formal register that detectors often misread. When a single percentage becomes a verdict, the conversation ends before anyone has looked at the lines that actually raised the flag.
I also wanted similarity and rewrite tools that stay honest about what they are. A match on the public web is a reason to cite or revise, not a finding of misconduct. A humanizer that loosens stiff phrasing is a draft to review, not a promise that the next check will look human.
The name comes from the flyleaf, the blank page at the front of a book. Flyleaf is that first look: useful, bounded, and not the whole story.
How I Put That Idea Into the Product
Flyleaf is a Flask web app. Checking a draft is free. You paste text or upload a file on one page, then choose Check for AI, Similarity, or Humanize. The AI checker uses a public DeBERTa model already trained on human and AI prose. Similarity embeds passages with MiniLM and compares them with public pages found for the draft. Humanize runs a smaller T5 rewrite model so the meaning should stay while the wording loosens.
"A percentage is a reason to reread a passage. It does not decide that someone used a chatbot."
The interface keeps three cues separate on purpose. The number in the pie is only the AI-like share. Inconclusive lines sit beside it and are not folded into that number. Confidence drops on short text, mixed sections, and writing that may not be English. Under about three hundred words, the result is a sketch, not a report you should treat as final.
For a text-based PDF, the download is the same reading you see on screen: a cover with the headline score, an overview of the groups, then the original pages with cyan marks on AI-like lines and purple on inconclusive ones. Sign in if you want that report kept with an account. A check you do not keep lasts only while the app is running.
What is stored, what is hashed, and how Google sign-in is linked is spelled out on the privacy note.
What You Do With It
- Open Flyleaf and paste a draft, or upload a text-based PDF or Word file.
- Run Check for AI to see AI-like, inconclusive, and human-like shares side by side.
- Optionally run Similarity against public pages, or Humanize for a plainer draft you still review yourself.
- Download or print the PDF report when you want the marked pages to travel with the conversation.
The Report You Keep
Pasted text stays on the page. A PDF opens as a short report so a teacher, classmate, or editor can walk the same highlights you saw, instead of arguing over a lone percentage.
Who It Is For
Students can check a draft before submission and see which lines a reviewer might pause on. Teachers can walk into a conversation with the highlighted page rather than with a single number. Writers can look at a passage edited with a model and keep a PDF if they want a record of that look.
How You Can Build Something Like This
If you want a portfolio tool in the same spirit, start with a narrow promise and keep the model from becoming the product story. Flyleaf’s stack is deliberately ordinary: Python, Flask, PyTorch, Hugging Face Transformers, and a small amount of front-end JavaScript. The hard part is not the framework. The hard part is deciding what the score is allowed to mean.
- Pick one honest job. Mine was “a careful first look,” not “proof of AI use.” Write that sentence before you train or download anything. Every screen should still make sense if someone quotes that sentence back at you.
- Use a published model, then wrap it in your own rules. Flyleaf does not train a detector from scratch. It loads desklib’s AI-text detector, MiniLM for embeddings, and a compact T5 humanizer. Your contribution is the product: sentence grouping, separate inconclusive marks, PDF layout, status polling while weights load, and copy that refuses to overclaim.
- Ship the interface people will actually argue from. A pie chart alone invites myth. A marked page invites reading. Build the report early, even if the first version is plain HTML and print CSS.
- Plan for a weak host. Shared cPanel or Passenger will kill large models, thread pools, and Hugging Face downloads. Prefer smaller weights, load them offline from disk, keep one worker process, and surface short errors when a tool fails instead of freezing the whole page.
- Separate similarity, detection, and rewrite. They answer different questions. Mixing them into one “AI score” teaches the wrong lesson. Let each mode fail or succeed on its own.
- Publish what you store. A free checker still needs a privacy note: what lives in the browser, what hits the server, what an account keeps, and what a password hash is. Trust is part of the design.
You do not need a research lab to start. You need a clear boundary, a model you can run on the hardware you actually have, and a UI that makes the uncertainty visible. If you build a sibling tool, keep the estimate labeled as an estimate. That restraint is what makes the project useful.
What I Left Out on Purpose
There is no claim of perfect accuracy. There is no paid wall in front of a basic check. Humanize is not sold as a way to beat a detector. Similarity is not sold as a misconduct finding. I left those promises out because they are the ones that make writing tools hard to trust.
Flyleaf is free to use and it is a portfolio project by Tanui Kipngetich Sila. It will not decide a case for you. It will help you see which lines deserve another reading, in the language of a first look rather than a final judgment.
Building something nearby? Start from the same idea: one honest job, published models, and a report people can read together. Open the app, then shape your own wrapper around the question you actually want to answer.
Leave a comment
Email stays private. Name and email are optional.