KRZResumeBuilder — Resume & Cover Letter Optimizer

Create your professional profile once, then tailor your resume and cover letter to every job. Built on the ApplyAiJobBoard design system (tokens in src/styles/tokens.css).

The app follows the product flow end to end:

Landing page → Sign up / sign in → Create professional profile → AI creates Master Resume + Master Cover Letter → Dashboard → Tailor resume to a job (upload or paste) → Analyze Job → Job Match Score → Create Application → Review & edit → Download PDF / Word / both → user applies manually.

With no configuration it runs on an in-memory store, so no database or API keys are needed to try it. Set MONGODB_URI and it stores everything in MongoDB instead (see Deploy).

Quick start

npm install
npm run dev            # http://localhost:3000
npm test               # ~100 tests: credits, fabrication guard, full flow, accounts, profile, uploads, both data stores

Create an account on the sign-up page. The in-memory store is empty after every restart, so you set up your profile again each time.

Sign-in

Email + password works out of the box (scrypt-hashed). Google and LinkedIn buttons appear automatically when AUTH_GOOGLE_ID/SECRET and AUTH_LINKEDIN_ID/SECRET are set. Set AUTH_SECRET in production.

Using a real model

Add ANTHROPIC_API_KEY (and/or OPENAI_API_KEY) to .env.local. Each workflow's model is set in src/server/ai/routing.ts. With no key, every workflow falls back to a deterministic mock provider.

Data storage

Setting	Store	Data
nothing set	in-memory store	resets when the server restarts
MONGODB_URI	MongoDB (Atlas)	persists

Both implement the same AppStore interface (src/server/store/types.ts), and one shared test suite (tests/store-contract.ts) runs against both. prisma/schema.prisma is a PostgreSQL model for a possible later phase and is not used today.

The MongoDB tests start a throwaway local MongoDB with mongodb-memory-server. The first run downloads an official MongoDB server binary (a few hundred MB, cached afterwards); without network they are skipped with a warning (MONGODB_TEST_REQUIRED=1 makes that an error). To test against your own server instead, set MONGODB_TEST_URI; each test uses a throwaway database named resumeai_test_<random> and drops it.

Deploy (Vercel + MongoDB Atlas)

The in-memory store cannot work on Vercel (every serverless instance would have its own empty memory), so a deployment needs MongoDB.

1. MongoDB Atlas

1. Create a free M0 cluster at https://www.mongodb.com/atlas, in a region near your Vercel region.
2. Database Access → add a database user with a strong password (read and write to any database).
3. Network Access → allow 0.0.0.0/0. Vercel has no fixed IP addresses. (Or install the MongoDB Atlas integration from the Vercel Marketplace, which manages this and sets the connection variable for you.)
4. Connect → Drivers → copy the connection string and put the database name before the ?: mongodb+srv://USER:[email protected]/resumeai?retryWrites=true&w=majority (URL-encode special characters in the password.)

You do not create collections or indexes by hand: on the first request the app creates its indexes and seeds the credit prices (existing prices are never overwritten). Collections: users, profiles, cv_imports, jobs, analyses, matches, ats_reports, resumes, cover_letters, ai_runs, credit_ledger, credit_balances, credit_actions. Atlas is always a replica set, which the credit ledger needs (it writes balance and entry in one transaction).

2. Vercel

1. https://vercel.com → Add New → Project → import this repository. It is detected as Next.js.
2. Add Environment Variables (Production, and Preview if you use it):

Variable	Required	Notes
MONGODB_URI	yes	the connection string from step 1
AUTH_SECRET	yes	production refuses to start without it: openssl rand -base64 32
MONGODB_DB	no	overrides the database name in the URI
ANTHROPIC_API_KEY	yes, in production	the AI that reads CVs and writes resumes and letters (each workflow's model is in src/server/ai/routing.ts). Locally, without a key the built-in sample writer is used. In production a missing key switches AI actions off (people see "AI writing isn't switched on for this site yet" and nothing is charged); /api/health says so
ANTHROPIC_WORKSPACE_ID	no	only if Anthropic answers "This API key is not scoped to a workspace": the workspace ID (starts with wrkspc_), sent with every request. Not needed for a key created inside a workspace
AI_ALLOW_MOCK	no	true lets a production preview use the sample writer. Never set it on the real site
AI_FORCE_MOCK	no	true forces the sample writer even when keys are set (tests)
CONTACT_EMAIL	recommended	the address shown on the Privacy policy and Terms pages
ALLOW_MEMORY_STORE	no	true lets a Vercel preview run on the temporary in-memory store (data is lost). Leave unset
AUTH_GOOGLE_ID/SECRET, AUTH_LINKEDIN_ID/SECRET	no	enable social sign-in
RESEND_API_KEY, EMAIL_FROM	for "Forgot password?"	send reset emails through Resend. EMAIL_FROM must be on a domain you verified there (e.g. KRZResumeBuilder <[email protected]>); Resend's test sender only reaches your own address. Without a key people cannot reset a forgotten password; /api/health shows

3. Deploy. Every push to main redeploys.

Never commit credentials: .env*.local, *.env and *credentials* are git-ignored; put secrets in Vercel's settings.

Good to know

• Is the site set up? Open /api/health on the deployed site: it says (yes or no, never the values) whether AUTH_SECRET and MONGODB_URI are set and whether the database answers, and explains each problem in a sentence. Sign-up and sign-in show the same message instead of a raw "server configuration" error.
• Upload size: Vercel serverless functions accept request bodies of about 4.5 MB, so CV and job-description uploads are limited to 4 MB.
• AI time limit: pages and actions are set to a 60 second limit (maxDuration in src/app/(app)/layout.tsx), downloads to 30 seconds. Raise it there if a real model needs longer (your Vercel plan sets the ceiling).
• Forgot MONGODB_URI? On Vercel the app refuses to start without it (instead of silently using temporary memory) and shows the reason in the function logs. ALLOW_MEMORY_STORE=true overrides this for throwaway previews.
• Backups: the free Atlas tier has no automated backups. Take a manual export before anything risky.

How it's organized

src/app/            routes only (pages compose components, actions stay thin)
src/components/     ui/ (design-system kit) · layout/ · features/
src/server/
  ai/               Career AI Agent: providers/, workflows/, guard/, runner, scoring, routing, pricing
  credits/          ledger service + memory, MongoDB and (unused) Prisma stores
  documents/        ResumeDocument model, templates/classic (PDF + DOCX), export
  services/         job-pipeline — the business logic for each step
  store/            AppStore interface + in-memory and MongoDB implementations; db/ holds the MongoDB connection
  queries/          page read-models
  auth/ billing/
prisma/             schema + seed
tests/

Key decisions

• The AI never writes facts. Workflows return reworded text plus evidence ids; code copies employers, titles, dates, schools and certifications from the profile (assembleResume).
• Fabrication guard in code, not just prompts (server/ai/guard): rejects unknown evidence ids, numbers not in the cited evidence, skills not in the profile, achievements borrowed across roles, headlines never held, and unknown organizations. One retry with the violations, then refund.
• Gap questions, not guesses. Missing requirements become Yes / No / Skip questions. A Yes stores the user's own words as an approved fact — the only new information AI may use.
• Scores are computed, not generated. The model classifies each requirement; fixed weights turn that into Job Match and ATS Compatibility scores, so they're stable and explainable.
• Credits are a ledger. Reserve before the call, settle or refund after; prices are rows in CreditAction, so changing a price is not a deploy. Every model call is an AIRun with tokens and USD cost.
• Provider abstraction. One AIProvider interface; Anthropic, OpenAI and Mock adapters.
• Versions are immutable. "Restore" copies an old version forward.

Credits per action (editable in CreditAction)

Analyze Job 10 (analysis 5 + match 5) · Create Application 25 (resume 15 + cover letter 10) · ATS check 3 · cover-letter rewrite 3 · Master documents free. New accounts get 60 credits.

Next

• PrismaAppStore (persistent accounts and documents), file storage for uploads
• Stripe plans and credit packs, admin pricing page
• Rate limits, prompt evaluation set against real job descriptions
• More resume templates