Glyph: I Built an AI Writing Assistant, Almost Turned It Into a Startup, Then Abandoned It for 6 Months — Here's How I Finally Finished It

This is a submission for the GitHub Finish-Up-A-Thon Challenge

What I Built

I write a lot. And like most people who write a lot, I also use AI constantly to fact-check, generate ideas, break through writer's block, understand something faster, or honestly, just be lazy and let it do the heavy lifting. The problem was always the same, my editor was in one tab, the AI was in another, and I was copy-pasting between them like it was 2012.
So I built Glyph, an AI-powered writing assistant that lives right inside your editor. No tab switching. No copy-pasting. Just you, your document, and an AI that can generate content from scratch, edit what you've already written, or answer questions about your text, all in the same window. It also supports video transcription, paste a link to any online video (not YouTube) and Glyph will pull the transcript directly into your editor.
The idea came from a real problem I faced personally, and one I watched non-technical colleagues struggle with daily at a previous job. They were spending significant time manually doing tasks that a tool like this could handle in seconds. Glyph was built to close that gap.
I originally built it as an MVP for the Mux Hackathon challenge. I genuinely believed in it enough to start thinking about it as a startup. Then life happened. No funding, no time, and Glyph quietly collected dust for six months.

Demo

Live Url: Live Demo
Test Login Email: [e-mail]
Test Login Password: GlyphTest2026
Repository: Github Repo

The editor:

Generate mode — creating content from scratch:
You type a prompt in the AI panel on the right, hit Send to AI, and the content appears directly in your document. No copy-pasting.


Ask mode — questions about your text:
Switch to Ask, type your question, and the AI responds in the panel on the right, leaving your document untouched.


Edit mode — rewriting selected text:
Switch to Edit, type your instruction, and the AI rewrites your entire document based on what you asked. You can also select a specific section ,notice the context chip in the bottom right of the input changes from "Whole doc" to the word range (e.g. w15-w68) showing the AI knows exactly what you've highlighted.



Video transcription:
Click the 📎 attachment button in the AI panel, paste any direct video URL (supports .mp4, .mov, Vimeo, Dropbox, Google Drive, not YouTube), and Glyph transcribes the audio and drops the full transcript into your editor.

The Comeback Story

Six months ago, I zipped up the project folder, told myself I'd come back to it "soon," and moved on. I didn't.
When I finally reopened Glyph two days before this submission deadline, here's what greeted me:
The original UI:


Then I tried to run it locally:

The app wouldn't even start. After six months, the Groq SDK had updated and was now incompatible with the version of httpx pip installed. The deployed version was frozen in time and still worked, but locally, it was dead on arrival. The fix was one command:

pip install --upgrade groq httpx

But it was a perfect metaphor for the whole project, functional from the outside, quietly broken underneath.
What the code looked like inside
Opening the project was like reading a letter from a past version of myself who had too much caffeine and too little sleep. The highlights:

A random junk comment in views.py: #nsfkjhr;igeorgljherighj;rile. I have genuinely no memory of typing this
The Groq client initialized twice, once at the module level doing nothing, and again inside the function that actually needed it
A duplicate handleVideoSubmit function, one empty ghost declaration, one real implementation sitting right below it, both with the same name

What I changed
Bugs fixed:

Removed the duplicate handleVideoSubmit declaration
Deleted the dead Groq client at the module level
Cleared out the junk comment
Fixed the dependency conflict and updated requirements.txt

UI redesigned:

Login page: purple gradient → clean white, confident dark button, proper breathing room
Editor header: cramped gray user box → clean pill with avatar initial
AI panel: floating dead-space icon → tight layout, segmented pill control for Generate/Edit/Ask
Sidebar: heavy blue borders → subtle hover states, Notion-style document list

Before and after:

My Experience with GitHub Copilot

I want to be honest here because I think the real story is more useful than a highlight reel.
Understanding code I'd forgotten
The first thing I did was open Copilot Chat in VS Code and ask:

@workspace explain what this project does and how it's structured

Copilot reviewed my files and gave me a clean breakdown in seconds. Django app, AI endpoints, Mux video transcription, Supabase logging, cookie-based auth. In five minutes I was back up to speed on a codebase I hadn't touched in six months. That alone saved me hours of re-reading my own code.
Then I pushed further:

@workspace where is the messiest code in this project?

It flagged the junk comment, the duplicate Groq client, the repeated Supabase logging boilerplate, and the manual CORS headers scattered everywhere. It was like asking a senior engineer for a brutally honest code review.
Redesigning the login page
I gave Copilot a specific prompt, replace the gradient, clean up the logo, make the button dark, keep all HTML and JS untouched:


This worked well. Scoped, specific, clean output.
The failed editor redesign — and what I learned
Here's where I'll be straight with you: Copilot didn't always get it right.
When I asked it to redesign the entire editor in one prompt, it generated a response, applied changes and broke everything. Elements scattered, alignment gone, the whole UI worse than before.

I caught it before shipping, restored my backup, and changed approach entirely. Instead of one massive prompt, I went surgical — one CSS section at a time, reviewing each change in the browser before accepting. Copilot's lesson here: it's a powerful tool, but it needs precision and a human checking the output. It is not a replacement for judgment.
Fixing the duplicate function

@workspace I have a duplicate function declaration for handleVideoSubmit around line 1367 one is empty and one has the real implementation. Can you explain why this is a bug and fix it?

Clean fix, clear explanation. Copilot at its best. Specific problem, precise ask, useful answer.

Glyph isn't finished-finished. Save and Export are still stubs. Real user authentication is on the roadmap. The startup dream isn't completely dead.
But it runs. It looks good. The AI actually works. And after six months sitting abandoned, it finally shipped.
That's enough for today.