Applywise — AI Job Match Copilot

Upload your resume once, then click a button on any job posting to get a match score, missing skills, ATS notes, salary info (converted to BDT), and AI-generated interview questions — all computed in your own browser with your own AI API key. Nothing is uploaded to any Applywise-owned server; there isn't one.

Live app: applywise-copilot.vercel.app

This repo has two parts that work together:

/ (root) — the web app (Setup + Results pages), a static React site
/extension — the Chrome/Edge/Brave/Firefox extension that scrapes job postings and talks to the AI provider

Demo

demo.mp4

A complete walkthrough of how Applywise works.

Quick start

This gets you from nothing to analyzing your first job posting. Takes about 5 minutes. No coding, no cloning, no installing Node — everything below happens in the browser.

1. Open the app and download the extension

Go to applywise-copilot.vercel.app. If the extension isn't detected in your browser yet, you'll see an "Install the browser extension" card at the top of the Setup page — click Download extension (.zip). (You can also grab it directly at applywise-copilot.vercel.app/applywise-extension.zip.)

This zip is always freshly built from whatever is currently deployed, so you always get the latest version.

2. Load it into your browser

Chrome / Edge / Brave:

Unzip the downloaded file — you'll get a folder named applywise-extension
Open chrome://extensions (or edge://extensions, brave://extensions)
Toggle Developer mode on (top-right corner)
Click Load unpacked and select the applywise-extension folder
You should see the Applywise icon appear — pin it via the puzzle-piece icon in your toolbar for easy access

Firefox: see Browser support below — it's a slightly different flow (about:debugging, select manifest.json directly).

3. Get a free Gemini API key

Applywise defaults to Google's Gemini API, which has a free tier.

Go to aistudio.google.com/apikey and sign in with a Google account
Click "Create API key"
Copy the key — you'll paste it into Applywise in the next step

Prefer a different free provider? Several are built in — pick one in the next step instead of Gemini:

Groq — a genuinely free, no-card tier on very fast hardware (Llama 3.3 70B, GPT-OSS 120B/20B). Note this is not the same as Grok/xAI, which is paid.
Cerebras — 1M tokens/day free and extremely fast, but the free tier caps context at ~8k tokens, so a very long resume may not fit.
Mistral — a free "Experiment" tier covering all models, no card.
OpenRouter — one key reaches several free models (Nemotron 3, Tencent HY3, Gemma 4, GPT-OSS), though they're noticeably slower than Gemini's and share a daily cap.
Cohere — a free Trial key (1,000 calls/month, non-commercial).
Hugging Face — routed through Hugging Face's Inference Providers (GPT-OSS 120B, GLM-5.2, Qwen3 8B/14B/32B, Qwen3-4B-Instruct-2507). Free accounts get $0.10/month of credit — a small trial allowance, not an ongoing free tier like Gemini's.

4. Set up your resume and API key

Back on applywise-copilot.vercel.app (refresh the page if it was already open, so it picks up the extension):

Under Resumes, click Upload resume and pick a PDF (up to 3)
Under AI provider, leave it on Gemini and paste the API key from step 3
Click Save settings — you should see a checkmark confirm it saved

5. Try it

Go to any job posting (LinkedIn, bdjobs, Indeed, a company careers page, etc.) and look for the "Analyze with Applywise" button in the bottom-right corner of the page. Click it right away — no need to scroll down the page or click any "see more"/"show more" description toggle first, the extension handles that for you. After a few seconds a new tab opens with your match score, skills breakdown, and job details.

That's it. You're done with setup.

Day-to-day use

Multiple resumes: upload up to 3, each with a profile name (e.g. "SQA", "SWE"). If you have more than one saved, the extension asks which one to use each time you analyze a posting.
Turning it off: click the Applywise toolbar icon — there's an on/off toggle that hides the floating button on every page until you turn it back on.
Interview prep: on the results page, click "Generate interview questions" for up to 20 likely questions with suggested answers, generated on demand (not part of the initial analysis, so you don't pay the extra API call unless you want it).
Session history: past analyses stay listed in the sidebar of the results page — local to this browser only. Remove one entry at a time or clear the whole list with a confirm step.
Switching AI providers: on the Setup page, pick from Gemini, Groq, Cerebras, OpenRouter, Mistral, Cohere, Hugging Face, DeepSeek, GLM, OpenAI, Anthropic, or Grok (xAI) — free-tier/trial options are listed first. Each has its own API key and model list; you can add multiple custom model IDs per provider (not just one) for anything not in the preset list. The API key field links straight to that provider's official key-creation page, so you're never stuck guessing where to go. Each provider remembers its own API key — switch providers and back, and the key you already saved for it is still there, no re-entering required. (Note: Groq at groq.com is a different provider from Grok (xAI) — the former has a free tier, the latter is paid.)
Real-time model availability check: the Setup page checks your selected provider's live model catalogue (fetched through the extension using your saved key, cached ~1 hour) and flags any preset or custom model that's no longer offered — both inline in the dropdown ("⚠ not in live list") and as a note under it, so you find out a model was retired before an analysis fails on a 404. Most providers only reveal retirement the moment a model actually drops from their catalogue, but OpenRouter and Mistral both publish an upcoming retirement date per model — when your selected model has one, it's flagged as "⚠ going away <date>" (in the dropdown and the status line) while it's still working today, so you can switch before it actually breaks instead of after. Needs the extension installed and a saved key.
Auto-fallback to other free models: a single toggle in the AI provider card. When it's on and your selected model times out or hits a rate limit / high-demand error, Applywise works down the free models — strongest first, fastest last — until one answers, instead of failing the analysis. Turn it off to analyze with only the model you picked, which is what you want when comparing models: a silent hand-off would credit another model's output to the one you selected. Your saved model is never changed either way; the results page shows which model actually answered.

It applies to the free-tier/trial-credit providers with several models to fall back to — Gemini, Groq, Cerebras, OpenRouter, Mistral, Cohere, and Hugging Face — and stays inert for the paid/single-key ones. You'll see it happen live: the floating "Analyze" button updates in real time ("gemini-3-flash-preview was busy — retrying with gemini-flash-latest…") instead of sitting on "Analyzing…" the whole time.

Note that OpenRouter's free models are much slower than Gemini's (measured: ~12s to ~200s per analysis, depending on the model and how loaded its host is), so expect to wait.
Extension update checks: "Load unpacked" installs can't auto-update (only Chrome Web Store/Firefox Add-ons installs can), so the extension popup checks the latest deployed version on every open and shows a banner — with a matching one on the Setup page — if yours is behind, so you're never left guessing whether a re-download is needed after a push. The Setup page's banner also lists a short "What's new" changelog for every version between yours and the latest, fetched fresh on every page load (public/changelog.json) so it reflects a just-pushed release immediately.

Browser support

Browser	Status
Chrome / Edge / Brave	Works out of the box on Windows, macOS, and Linux — same steps above, since they all run the same Chromium extension engine regardless of host OS.
Firefox	Supported (Firefox 121+). See below.
Safari	Requires converting the extension via Xcode on a Mac — not something that can be done from the downloaded zip directly. See below.

Firefox

Unzip the downloaded extension as in Quick start
Open about:debugging#/runtime/this-firefox
Click Load Temporary Add-on…
Select manifest.json inside the unzipped applywise-extension folder directly (not the folder itself — Firefox wants the manifest file)

Note: Firefox's "temporary add-on" loading resets when you close Firefox — you'll need to reload it each session. For a permanent install, the extension would need to be signed by Mozilla (via addons.mozilla.org), which is a separate publishing step not covered here.

Safari

Safari doesn't load unpacked web extensions the way Chromium/Firefox do — it requires packaging as a native Safari Web Extension via Xcode. If you're on a Mac and want to try it:

Install Xcode (from the Mac App Store)
Unzip the downloaded extension, then from that folder run:
```
xcrun safari-web-extension-converter applywise-extension
```
This generates an Xcode project — open it in Xcode and build/run
Enable the extension in Safari's Settings → Extensions (you may need to enable "Allow Unsigned Extensions" in Safari's Developer menu for local testing)

This step can't be completed from a non-Mac environment, so it hasn't been tested as part of this project — treat it as a starting point.

Features

Resume matching — match score (0–100), matching/missing skills, missing ATS keywords, ATS parseability notes, and concrete improvement suggestions. Scoring follows an explicit enumerate-then-compute rubric (required/ preferred skill coverage + experience fit, weighted and combined by a fixed formula) rather than a freeform "impression" score, so re-analyzing the same resume against the same posting lands close to the same score instead of swinging wildly between runs. The result leads with a qualitative Strong/Partial/Needs work verdict rather than the raw number — verified live across providers, different AI models can land on a somewhat different exact score for the same resume/posting pair (they judge borderline requirement coverage a little differently), but the verdict band stays consistent, so that's the number worth trusting at a glance. Skill matching is meaning-based, not string-based: the same technology written differently between your resume and the posting (an acronym vs. its spelled-out form, a different capitalization) still counts as a match, with the wording difference itself flagged as an ATS note since a real ATS keyword filter does match literally. ATS notes and suggestions are justified and highlight the specific skills/keywords they're about, instead of reading as an undifferentiated block of text. The prompt explicitly forbids naming any skill or technology that isn't actually present in your resume or the posting, so fast/lite models can't invent a finding (e.g. claiming a term appears in your resume when it doesn't) — a real defect a small model can otherwise fall into by copying an illustrative example instead of reasoning about your actual documents.
Robust job posting extraction, no scrolling or clicking required — the title, company, location, and full description are read from the page with several safeguards: before reading, the extension hydrates any lazily- rendered sections and clicks in-place "see more"/"show more" expanders on your behalf, then waits for the DOM to settle, so the full description is captured on the very first click — you never need to manually scroll the page or expand anything yourself first. Text structure (<br>/<li>/<p> boundaries) is preserved so labels stay attached to their values, page chrome (site footers/nav bars swept in by a wide container) is stripped before the text ever reaches the AI, the largest real description block is chosen (not whichever markup comes first), JSON-LD and DOM sources are merged for the richest result, and — importantly for single-page-app boards like LinkedIn — the posting is re-read at the moment you click Analyze (not when the button first appears), so a half-loaded page or list preview never gets analyzed by mistake. Nothing keys off hashed CSS class names, which churn every deploy.
Job details extraction — employment type, work mode (remote/hybrid/ onsite), location, company, salary, required experience (e.g. "3-5 years", "Entry level"), and a benefits/perks list (provident fund, gratuity, gym membership, parental leave, festival/performance bonus, flexible hours, medical allowance, and the like), read directly from the posting. Works across languages (a posting in Japanese or Bengali still gets a readable English location), across boards with non-standard markup (skill chip widgets, overview boxes rendered outside the main description), and picks the AI's reading over placeholder text like "Unknown company" when the page's structured data is missing. Anything the posting genuinely doesn't state shows as "Not available" rather than being guessed.
Live currency conversion — non-BDT salaries are converted to BDT using live exchange rates, shown alongside the original figure with the correct per-year/per-month/per-hour period; skipped entirely when a salary has no actual numbers (e.g. "Negotiable").
Multi-provider AI support — Gemini, Groq, Cerebras, OpenRouter, Mistral, Cohere, Hugging Face, DeepSeek, GLM (Zhipu/Z.ai), OpenAI, Anthropic, and Grok (xAI), with free-tier/trial-credit options listed first, multiple custom-model-ID slots per provider as an escape hatch, and a direct link to each provider's own key-creation page next to the API key field. (Groq at groq.com is a distinct provider from Grok/xAI — the former is free, the latter paid. Hugging Face's free credit is a small $0.10/month trial allowance, not an ongoing free tier.)
Real-time model availability check — the Setup page validates each preset and custom model against the provider's live /models catalogue (fetched via the extension with your saved key, cached ~1 hour) and flags retired ones in the dropdown and a status line, so a model that's been pulled is caught before it 404s an analysis. For OpenRouter and Mistral, which both publish an upcoming retirement date per model, a still-working model due to go away soon is flagged too ("⚠ going away <date>"), ahead of the actual cutoff
Auto-fallback across free models, with live progress — one toggle in the AI provider card; when on, timeouts, rate limits, high-demand errors, and responses that come back unusable all move the analysis to the next model in the same provider (strongest first, shown in real time on the floating "Analyze" button) rather than failing it. Works for the providers with several models to fall back to (Gemini, Groq, Cerebras, OpenRouter, Mistral, Cohere, Hugging Face). Errors another model can't fix — a bad API key, say — still surface immediately instead of burning the chain. The model that actually answered is shown on the results page. Turn it off to pin an analysis to exactly one model
Interview prep — up to 20 likely interview questions with suggested answers, generated on demand and cached so revisiting the page doesn't re-spend an API call
Multiple resume profiles — up to 3 resumes, each labeled (e.g. "SQA", "SWE"); picked per-analysis when more than one is saved
Session history & status tracking — past analyses stay listed locally, with a lightweight Saved/Applied/Interviewing/Offer/Rejected status per job, and smooth remove/clear-all animations
On/off toggle — turn the floating "Analyze" button off globally from the extension popup
One-click extension download + update checks — the Setup page detects whether the extension is installed and, if not, offers an always-current download built fresh on every deploy; if it's installed but out of date, both the popup and the Setup page flag it (unpacked installs can't auto-update the way store-installed ones do)
Cross-browser — Chrome, Edge, Brave, and Firefox, via a shared browser.*/chrome.* compatibility layer

Deferred (not in this build)

Chrome Web Store / Firefox Add-ons store listing — the only way to get a true one-click install and genuine silent auto-updates, but requires a developer account (Chrome: $5 one-time fee) and a review process. Deliberately not pursued right now; the download-zip + "Load unpacked" flow plus the update-check banner (see Features) is the stand-in until that changes.
Cover letter generator
Manual per-site scraping overrides beyond the generic extractor
Cross-device history sync (would need accounts + a backend — contradicts the "no server" design)
Automated test suite / CI

Tech stack

React 19 + TypeScript — both the web app and the extension's UI
Vite — web app dev server/build
Tailwind CSS v4 — styling, with a dark/light theme via CSS variables
react-router (HashRouter) — Setup/Results page routing in the web app
esbuild — extension build (extension/build.mjs), separate from Vite
Chrome Manifest V3 — extension platform, via a browser.*/chrome.* compatibility shim for Firefox
Zod — runtime schema validation for AI provider responses

Architecture

No database, no accounts, no backend. Resume text, API keys, and job history all live in chrome.storage.local (the extension) or localStorage (the web app, when used standalone without the extension installed).
The extension does all AI calls. Both the initial match analysis and the "Generate interview questions" follow-up run in the extension's background service worker — a privileged context, not subject to a webpage's CORS restrictions. The web app never calls the AI provider directly.
Web app ⇄ extension bridge. The web app runs on its own origin and has no direct access to chrome.storage (that's extension-only). A second content script (app-bridge.js), injected only on the web app's own origin, relays window.postMessage calls from the page to chrome.storage/chrome.runtime. This is a deliberately narrow trust boundary — see Local development if you're pointing this at a different deployed URL.
Cross-browser API shim. extension/src/lib/browser-api.ts prefers the native browser.* API (Firefox) and falls back to chrome.* (Chrome/Edge/ Brave), so the same source works on both without behavior changes on Chromium browsers.
Extension packaging. scripts/package-extension.mjs builds the extension and zips it into public/applywise-extension.zip as part of npm run build, so every deploy (and thus every download from the Setup page) ships the current source, never a stale snapshot. It also writes public/extension-version.json from extension/manifest.json's own version field, which is what the popup checks against on every open (see extension/README.md).
One-way progress channel. While an analysis runs, the background worker pushes ANALYZE_PROGRESS messages to the originating tab (which model is being tried, whether a Gemini fallback kicked in) — a plain runtime.sendMessage/onMessage push with no reply expected, separate from the request/response messages (ANALYZE, GET_RESUMES, etc.) used everywhere else.

Project structure

ApplyWise/
├── src/                          # Web app (Setup + Results pages)
├── extension/
│   ├── src/
│   │   ├── background/           # Service worker — AI calls, storage, tab opening
│   │   ├── content/               # content-script.js (the floating button) +
│   │   │                          # app-bridge.js (web app ⇄ extension storage bridge)
│   │   ├── popup/                 # Toolbar popup UI
│   │   └── lib/                   # Shared: types, AI clients, extraction, storage
│   ├── manifest.json
│   └── build.mjs                  # esbuild build script
├── scripts/
│   └── package-extension.mjs      # Builds + zips the extension for the Setup page download
└── applywise-project-plan.md

Local development (running from source)

Only needed if you're contributing to Applywise itself, or want to run your own deployment instead of using the hosted one. Everyday users can stop at Quick start above.

Prerequisites

Node.js 20 or newer — check with node -v. Get it from nodejs.org if you don't have it.

Get the code and install dependencies

git clone https://github.com/jhrahman/applywise.git
cd applywise
npm install
cd extension && npm install && cd ..

Run the web app

npm run dev

Starts the web app at http://localhost:5173.

Build and load the extension

cd extension
npm run build     # one-off build to extension/dist
npm run dev       # or: watch mode, rebuilds on save
cd ..

Load extension/dist as an unpacked extension (same steps as Quick start step 2, but point at extension/dist instead of an unzipped download). By default it points at http://localhost:5173 — see extension/src/lib/config.ts.

After any extension code change, click the reload icon on the Applywise card at chrome://extensions, then refresh any already-open tabs (an already-open tab's content script won't pick up the reload automatically).

Deploying your own copy (Vercel)

The web app is a static site — no server runtime needed, and no environment variables — every user's AI API key lives in their own browser, never on the server.

Push your fork to GitHub
In Vercel, click New Project and import the repo
Framework preset: Vite. Build command: npm run build (this also builds and zips the extension — see Architecture). Output directory: dist.
Deploy — Vercel auto-deploys on every push to your main branch after this initial setup

After deploying, point the extension at your production URL instead of localhost:5173:

Edit extension/src/lib/config.ts — change APP_URL to your Vercel URL
Edit extension/manifest.json — add your Vercel URL to the content_scripts entry whose matches currently lists the localhost and applywise-copilot.vercel.app URLs (this is what lets the Setup page talk to the extension's storage — see Architecture)
Rebuild the extension (cd extension && npm run build) and reload it at chrome://extensions

Privacy & security notes

Your AI provider API key lives in chrome.storage.local — per-browser- profile, not synced or encrypted beyond what Chrome/Firefox provide by default. That's normal for a personal tool, but worth knowing.
Job descriptions are scraped, untrusted content. Prompts sent to the AI provider wall off scraped text with explicit delimiters and an instruction to treat it as data, not commands — basic prompt-injection hygiene.
No cross-device sync. Everything lives in one browser's local storage; switching browsers or computers starts fresh.

Troubleshooting

"Extension context invalidated" / blank error after clicking Analyze: you reloaded the extension while a job tab was already open. Refresh that tab.
429 rate limit error: your AI provider's free-tier quota was hit. With auto-fallback on, the providers with several models (Gemini, Groq, Cerebras, OpenRouter, Mistral, Cohere, Hugging Face) try their other models first — if you still see this, all of them are limited; wait a bit or try a different provider. Some free tiers (Cerebras, Cohere) cap requests per minute account-wide, so switching models won't help there — wait for the window to reset. On OpenRouter specifically, a 429 on one model usually just means that model's upstream host is busy (fallback routes around it), whereas every free model failing at once points at the shared daily cap on free models — check https://openrouter.ai/activity. For other providers, try a lighter model on the Setup page.
404 "model no longer available": providers retire model IDs over time. The Setup page's real-time model check flags this ahead of time ("⚠ not in live list") once you've saved a key — when you see it, pick another preset, or choose "＋ Add a custom model…" and paste a currently-valid ID from your provider's docs/dashboard (you can keep several per provider).
"Failed to load extension: Manifest file is missing or unreadable": you selected the wrong folder — after unzipping the download, load the applywise-extension folder itself, not its parent.
Setup page doesn't detect the extension after installing it: refresh the tab — the install-detection check runs once when the page loads.
Analysis says the description is missing / "only LinkedIn boilerplate" even though the job clearly has one: make sure you're on the job's own page (not a search list or preview pane). Clicking Analyze re-reads the posting, hydrates any lazily-rendered content, and clicks "see more"-style expanders automatically — you shouldn't need to scroll or expand anything yourself. If it still comes back thin on a specific posting, that page's markup may be unusual (or the page is behind a login the extension can't see through); the strongest model (gemini-3-flash-preview) also tends to handle sparse pages best.

Name		Name	Last commit message	Last commit date
Latest commit History 32 Commits
.github/workflows		.github/workflows
docs		docs
extension		extension
public		public
scripts		scripts
src		src
.gitignore		.gitignore
.oxlintrc.json		.oxlintrc.json
LICENSE		LICENSE
README.md		README.md
applywise-logo.svg		applywise-logo.svg
applywise-project-plan.md		applywise-project-plan.md
applywise-theme-preview.html		applywise-theme-preview.html
index.html		index.html
package-lock.json		package-lock.json
package.json		package.json
postcss.config.js		postcss.config.js
tsconfig.app.json		tsconfig.app.json
tsconfig.json		tsconfig.json
tsconfig.node.json		tsconfig.node.json
vite.config.ts		vite.config.ts

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Applywise — AI Job Match Copilot

Demo

Contents

Quick start

1. Open the app and download the extension

2. Load it into your browser

3. Get a free Gemini API key

4. Set up your resume and API key

5. Try it

Day-to-day use

Browser support

Firefox

Safari

Features

Deferred (not in this build)

Tech stack

Architecture

Project structure

Local development (running from source)

Prerequisites

Get the code and install dependencies

Run the web app

Build and load the extension

Deploying your own copy (Vercel)

Privacy & security notes

Troubleshooting

About

Releases

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

Applywise — AI Job Match Copilot

Demo

Contents

Quick start

1. Open the app and download the extension

2. Load it into your browser

3. Get a free Gemini API key

4. Set up your resume and API key

5. Try it

Day-to-day use

Browser support

Firefox

Safari

Features

Deferred (not in this build)

Tech stack

Architecture

Project structure

Local development (running from source)

Prerequisites

Get the code and install dependencies

Run the web app

Build and load the extension

Deploying your own copy (Vercel)

Privacy & security notes

Troubleshooting

About

Topics

Resources

Stars

Watchers

Forks

Releases

Contributors

Languages