SayResume
Fix your resume by talking to it. Upload a resume, say what to change, and watch the LaTeX source and the PDF update live. Pick Local (a model on your own Mac) or Cloud (a free hosted model) with one toggle.
Built for No Hands: The Wispr Challenge (WeMakeDevs × Wispr Flow).
- Live demo: https://sayresume.vercel.app (Cloud mode only; Local needs the app on your own computer)
- Demo video: add link
- Pitch (5 slides): open
pitch.html in a browser; arrow keys to move, Print → Save as PDF to export
The problem
Updating a LaTeX resume for each application is slow: move a section, reword a bullet, escape a % that silently comments out the rest of the line, recompile, repeat. Resume builders hide the source; LaTeX gives control but costs time. SayResume keeps the LaTeX and removes the typing.
What it does
| You say | What happens |
|---|
| "Make my Acme bullets sound stronger" | Rewrites those bullets, keeps the facts |
| "Add a bullet under Acme: cut page load time by 40 percent" | Adds \item … 40\% under the right job |
| "Move skills above experience" | Moves the section, nothing lost |
| "Add AWS and Docker to my skills" | Appends to the skills line |
| "Remove the projects section" | Deletes it cleanly |
| "Actually, undo that" | Instant undo (no model call) |
| "Add my job at Google" | Asks for title and dates instead of inventing them |
- Upload anything:
.tex loads as-is. .pdf, .docx, .txt, .md are read in your browser (PDF.js is bundled, so no third-party CDN) and converted to LaTeX by the model.
- Live view: changed lines flash in the source pane, a red/green diff appears in the log, the PDF recompiles in about 0.2 s (locally).
- Voice or keyboard: speak into the command box with Wispr Flow (hold Fn), macOS Dictation or anything else. Dictated text sends itself after a short pause; typed text sends on Enter. Nothing requires a particular dictation tool.
- Versions in your browser: the resume and its last 30 versions live in
localStorage; the server stores nothing.
Local vs Cloud
| Local | Cloud |
|---|
| Model | Qwen3.5-9B (4-bit) via oMLX on your Mac | Qwen 3.7 Flash via OpenRouter, with Mistral Small and Claude Haiku as fallbacks |
| Where your resume goes | nowhere: model and LaTeX compiler both run on your machine | the server (to compile) and OpenRouter (free models may log prompts) |
| Needs | Apple Silicon Mac, oMLX, tectonic | nothing; works on the hosted site |
| Speed | ~2–8 s per command | ~1–3 s per command |
| Cost | free | about $0.08 per 1,000 commands |
The toggle is in the header. On the hosted site Local is disabled: a cloud server can't reach your computer. Run the app yourself to use it.
How the understanding layer works
spoken command ─► model (local or cloud)
│ replies with small SEARCH/REPLACE blocks and MOVE lines
▼
apply to the .tex ─► safety checks ─► compile (tectonic) ─► new .tex + PDF
▲ │
└── error goes back ──┘ (self-corrects, up to 3 tries)
- Small edits, not rewrites. The model only emits the lines it changes, so it cannot quietly drop content. Each
SEARCH must match exactly one place (with whitespace tolerance for small-model drift).
- Moves are done by code. The model just says
MOVE: Skills BEFORE Experience.
- Guards. Rejected if a section is duplicated, disappears without being asked, or the LaTeX no longer compiles. The failure goes back to the model, which retries.
- No invented facts. The prompt forbids new employers, dates, degrees, metrics or skills, and tells the model to ask (
NOCHANGE: <question>) when a command is unclear.
- LaTeX gotchas handled:
40% becomes 40\% so it doesn't start a comment.
Run it locally
brew install tectonic # LaTeX compiler
omlx start # local model server (needs Qwen3.5-9B-4bit in ~/.omlx/models)
cp .env.example .env # optional: add OPENROUTER_API_KEY to enable Cloud locally too
npm start # http://localhost:3000
No npm dependencies; Node 20+. Get the model from mlx-community on Hugging Face (about 6 GB) into ~/.omlx/models/mlx-community/, then omlx restart. The first compile downloads TeX packages once (a few minutes), after that it is instant.
Deploy to Vercel
npx vercel link
printf '%s' "$OPENROUTER_API_KEY" | npx vercel env add OPENROUTER_API_KEY production --sensitive
npx vercel --prod
How it works there: the build step (scripts/prepare-tectonic.mjs) downloads the Linux tectonic binary and pre-fills its package cache; the functions in api/ copy that cache to /tmp on cold start and compile with it. public/ is served as static files. The server adds a per-IP rate limit; the OpenRouter key's own spending limit is the real cap.
Tests
npm test # edit engine + backends (model calls mocked, LaTeX compile is real) + DOCX extraction
demo-voice/ holds nine spoken test commands (macOS TTS, Indian-English voices, with filler words) used to check the voice path end to end: speech → local Whisper transcript → app.
Layout
lib/edits.mjs pure text logic: parse/apply edits, move sections, guards
lib/engine.mjs backends (local/cloud), compile, edit/import loops, HTTP handlers
api/*.mjs thin Vercel functions (edit, import, compile, health)
local.mjs local server: static files + the same handlers
scripts/ Vercel build step (tectonic binary + cache)
public/ UI (index.html, sayresume.js, extract.js, styles.css), sample.tex, vendored PDF.js
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
