SwiftApply

Inspiration

I'm a CS student, and last fall I watched the internship application season hit my friends like a truck. One guy applied to 127 companies. Another spent every weekend writing cover letters. The whole process had turned into a brutal numbers game.

What really got me was seeing someone spend 45 minutes writing a perfect cover letter, only to get auto-rejected three minutes later. The effort didn't match the reality. I kept thinking there had to be a better way.

When I found out about Chrome's built-in AI APIs, something clicked. What if AI could handle the repetitive stuff (reading job descriptions, writing cover letters, filling forms) while students focused on the parts that actually mattered? And what if it all ran locally so your data stayed private? That's where SwiftApply started.

What it does

SwiftApply is a Chrome Extension that helps you apply to jobs faster without sacrificing quality. It uses Gemini Nano to handle the tedious parts of applications.

When you open a LinkedIn job posting:

Job Analysis - It reads the entire description and pulls out the important stuff: key requirements, skills needed, salary info, and a quick summary. Takes about 3 seconds.
Profile Match - It checks your actual qualifications against what the job needs. Not fake percentages, but real criteria like "do you have the right degree?" and "do you have these skills?"
Why You Fit - Shows you specific reasons you're qualified, like "3 years experience in this field" or "currently working as X role."
Cover Letter Generator - Writes a personalized cover letter in 10-30 seconds using your actual work history and skills. No generic templates.
Application Tracker - Saves jobs with one click and lets you track status (Pending, Viewed, Interviewing, Rejected, Accepted). Includes stats and links back to the original postings.

Everything runs on your computer. Your resume and personal info never leave your device, there are no API costs, and it works offline after setup.

How we built it

Tech Stack

TypeScript and React for the frontend
WXT Framework for Chrome Extension development
Tailwind CSS with shadcn/ui components
Chrome Built-in AI (Gemini Nano) through the Prompt API

Architecture

I split SwiftApply into three main parts:

Content Script - Runs on LinkedIn pages and grabs job data (title, company, description, etc.). Uses MutationObserver to detect when you switch between jobs and debouncing to avoid running constantly.

Background Worker - Handles all the AI stuff. Manages Gemini Nano sessions, processes job analysis and cover letter requests, stores data in chrome.storage.local, and caches your profile for speed.

React Popup - The UI you see with three tabs (Home, Profile, Job Summarizer). Polls for updates to stay in sync and handles all the user interactions.

The AI Integration

Getting the Prompt API to work reliably took some trial and error (an entire week of debugging and reading docs 😭). I use structured outputs to make sure responses are always parseable:

// Structured output with schema constraints
const schema = {
  type: "object",
  required: ["cleanSummary", "salary", "skills", "requirements"],
  properties: {
    cleanSummary: { type: "string" },
    salary: { type: "string" },
    skills: { type: "array", items: { /* ... */ } },
    requirements: { type: "array", items: { type: "string" } }
  }
};

const result = await session.prompt(prompt, {responseConstraint: schema});

For cover letters, I give the AI your full work history, skills, and the job details. It generates the creative parts while my code handles formatting and structure. This keeps quality high and output consistent.

Profile Matching System

I didn't want to use AI-generated match scores because they felt misleading. Instead, I built a scoring system based on actual criteria:

$$\text{Profile Score} = 30(\text{Experience}) + 25(\text{Education}) + 15(\text{Location}) + 30\left(\frac{\text{Skills Matched}}{\text{Skills Required}}\right)$$

This gives honest scores users can trust and understand.

Challenges we ran into

1. Gemini Nano Availability

The biggest headache was dealing with model availability. Gemini Nano isn't always ready right away, it needs to download first. Had to add checks everywhere:

const availability = await LanguageModel.availability();
if (availability === 'after-download') {
  await LanguageModel.create(); // Triggers download
  // Show user a message to wait and try again
}

The API error messages weren't super helpful, so I added a lot of logging to figure out what was going wrong.

2. Getting Reliable AI Output

Early on, the AI would give me responses I couldn't parse. Sometimes valid JSON, sometimes creative prose, sometimes a mix. I learned to be extremely specific in prompts:

Tell it exactly what format you want
Use schema constraints every time
Include examples in the prompt
Always validate the output before using it

3. LinkedIn's Changing Structure

LinkedIn changes their page structure all the time, which broke my scraper constantly. Had to write flexible selectors that work across different layouts, add retry logic for slow-loading elements, and use MutationObserver carefully to catch changes without destroying performance.

4. State Management Across Extension Parts

Chrome Extensions have three separate contexts (background, content, popup) that can't directly share state. This was confusing at first. I ended up using chrome.runtime.messaging for communication, polling in the popup to stay updated, and caching important data in the background worker.

5. Cover Letter Quality

The first AI-generated cover letters were either too generic or too wordy. Took a lot of prompt tweaking to get them right:

Specified word counts (300-400 words)
Told it to avoid clichés
Gave it examples of what NOT to say
Passed in actual job descriptions from employment history

The breakthrough was including your full work experience so the AI had real examples to reference instead of making stuff up.

Accomplishments that we're proud of

It Actually Works

I built three different use cases with the Prompt API (job analysis, cover letters, profile matching) and they all work reliably. The code handles errors properly, manages sessions correctly, and performs well. This isn't a demo, it's something people can actually use.

Real Time Savings

SwiftApply genuinely saves time. Normal application: 50 minutes. With SwiftApply: 3 minutes. Over 100 applications that's 78 hours saved. The math is real.

Privacy First

Everything runs on your device. Your resume, work history, and personal info never leave your computer. Given how much data gets leaked these days, this actually matters.

Honest Scoring

I'm really proud that we don't use fake match percentages. The profile scoring system uses transparent criteria you can verify yourself. If it says you're missing a required skill, you are. If it says you have relevant experience, you do.

Quality Output

The cover letters are actually good. I've shown them to career counselors and consistently get 8-9 out of 10 ratings. They're personalized and professional without being obvious AI slop.

What we learned

About AI Development

On-device AI changes everything economically. With no API costs, I could build features that would bankrupt a startup using cloud AI. Students can generate unlimited cover letters without worrying about running out of credits.

Structured outputs are mandatory for production. You can't just cross your fingers and hope the AI gives you what you need. Schema constraints are the difference between a toy and a tool.

Context makes or breaks quality. Generic prompts get generic results. Including user context (skills, experience, location) transforms AI from "kind of helpful" to "actually useful."

About Chrome Extensions

WXT Framework is great. It brings modern web dev practices to extension development. TypeScript, React, hot reload, all of it just works.

Message passing needs careful planning. You can't wing the communication between content scripts, background workers, and popups. Need a clear protocol and state management strategy from the start.

Users care about permissions. I was careful to only request what I absolutely needed. People are rightfully suspicious of extensions these days.

About Building Products

Show your work early. I demoed SwiftApply to classmates constantly. Their feedback shaped major decisions, like ditching the fake skill percentages.

Solve your own problems. I built this because I needed it. That made every design decision obvious because I was the user.

Simple beats clever. I cut several "cool" features because they confused people. The simpler version is better.

What's next for SwiftApply

Short Term (Next 3 Months)

More Chrome AI APIs

Summarizer API for extremely long job descriptions
Rewriter API to adjust cover letter tone
Writer API for generating company-specific interest statements

More Platforms

Extend to Indeed, Glassdoor, and company career sites
The architecture is already modular, just need new scrapers

Better Tracking

Export to CSV/Excel
Email reminders for follow-ups
Calendar integration for interviews

Medium Term (6-12 Months)

Translator API

Apply to international jobs
Analyze postings in other languages

Interview Prep

Generate practice questions based on job requirements
AI role-play for technical and behavioral interviews

Analytics

Track success rates by company type
Identify patterns in rejections
See how you compare to others (anonymized)

Long Term Vision

Mobile App

Track applications anywhere
Get notifications on status changes
Edit profile on the go

University Integration

Partner with career services
Provide hiring trend data
Offer through university subscriptions

Broader Audience

Full-time jobs, not just internships
Non-tech fields
International markets

Community Features

Share company insights anonymously
Flag scam postings
Build a library of successful approaches

The Bigger Picture

SwiftApply started as a time-saving tool, but it could become something more important. Students at less prestigious schools, career changers without traditional backgrounds, and international students all face disadvantages in job searches.

AI assistance that's free and private could help level the playing field a bit. That's what I'm working towards.