This is a submission for the GitHub Finish-Up-A-Thon Challenge
Loom link: https://www.loom.com/share/d4ac87a0474f4a9eb10e4ebedad5b7a7
What I Built
Lynt is an AI-powered résumé and cover-letter builder with a visual editor, live print-accurate preview, one-click PDF export, and a public shareable page.
The core idea is not just generating text with AI — but letting AI apply structured edits directly into the document while preserving layout, formatting, and history. Users can rewrite bullets, reorder sections, and tailor résumés to job descriptions with full undo support.
The goal is to make editing a résumé feel faster and more reliable than copy-pasting between ChatGPT and a document editor.
It started as a hackathon project called ResumeForge, originally just a markdown → PDF tool. Over time, it evolved into a full SaaS with authentication, cloud storage, document ingestion (PDF/DOCX/images), an AI editing system, and a reliable PDF generation pipeline.
Demo
This project is currently in private beta while final stability and polish are being completed.
- Live: Private beta (not publicly available yet)
- GitHub: Private repository
- Screenshots: (see below)
Screenshots
- AI applying edits directly to a résumé
- Before vs after edit in editor
- PDF export matching live preview
- Upload → parsed résumé reconstruction flow
- Public résumé share page
The Comeback Story
Lynt began as a hackathon prototype built around a simple idea: markdown → PDF export.
It worked, but it was not reliable enough for real-world use.
Early state
The original version had clear limitations:
- AI suggestions could not be safely applied to documents
- PDF output often did not match the editor preview
- No production-grade authentication or storage layer
- No protection against data loss or inconsistent edits
It felt like a demo that “almost worked,” but not a product you could trust.
Turning it into a product
The focus shifted from adding features to improving reliability and correctness.
Instead of free-form AI output, the system was rebuilt around:
- Structured, deterministic document operations
- Validation layer for every AI-generated change
- Full undo/redo system using an action ledger
- Data-loss prevention rules (no silent deletions or overwrites)
- PDF parity between editor, export, and public view
A key shift was making the AI behave like an editor, not a generator.
Every change is:
- structured
- validated
- reversible
Production issues uncovered
When moved into real-world conditions, several issues surfaced:
- PDF rendering failures in serverless environments (
DOMMatrixissues) - SSE streaming instability on edge runtimes
- Public résumé pages breaking when underlying data was deleted
- Latency spikes caused by over-aggressive validation logic
These were not feature bugs — they were production reliability issues.
The outcome
The final system is not defined by features, but by predictability:
- AI edits no longer break layout
- No silent content loss
- PDF output matches the editor exactly
- Full document history with undo/redo
- Public pages remain stable and consistent
The biggest change was moving from “it works” to “it behaves reliably.”
My Experience with GitHub Copilot
Copilot helped mainly with accelerating repetitive development:
- API routes and boilerplate
- schema definitions
- UI scaffolding
- basic test generation
A Copilot coding agent was also used for a scoped feature (PR #21), which was reviewed and merged.
However, the core system design — especially the AI editing contract, validation system, and document safety model — required manual architecture decisions.
Final Notes
Lynt started as a hackathon experiment and evolved into a production-grade system focused on one goal: making AI-powered document editing reliable, deterministic, and safe.
The Finish-Up-A-Thon provided the push to complete the hardest part of any product — the reliability layer that turns a demo into something real.
Top comments (0)