Case study

Job Scout: a smart assistant for your job search

A complete web application that reviews new job offers every morning, scores them against your CV, drafts cover letters for the best ones and sends a summary by e-mail. AI pre-selects, a human decides.

Job Scout – offers scored against a CV, Microsoft and Google sign-in, morning e-mail and AI costs
Illustration of the app – companies and offers are fictional.
~3 days
of intensive work with an AI agent
2–3 months
estimate for a team of three without AI
600+
automated tests
21
recorded architecture decisions

Why it exists

Everyone knows job hunting: dozens of portals, hundreds of ads and hours of reading before you find the few offers that really fit. I wanted to show that with AI you can quickly build an application that does this work for you – while meeting everything expected of a real product: sign-in, multiple users, data protection, cost control and cloud operations.

It runs in production, I use it every day, and anyone can try it – without signing up, on a demo account with real current job offers.

How it works

  1. 1Sign-inWith a Microsoft or Google account – no new passwords.
  2. 2Your CVUpload a CV and a few notes (“fully remote, no Java”). AI builds your profile from them.
  3. 3Search every morningAt 6:30 only new offers are fetched; a cheap model triages titles, a stronger one scores the rest 1–10.
  4. 4Letter and e-mailCover letter drafts for the best offers, a morning summary by e-mail. Nothing is ever sent automatically.

Under the hood – for developers

For those who want to know how it is built. All decisions, including rejected alternatives, are recorded as ADRs in the public repository.

Sign-in and access

  • OpenID Connect with Microsoft and Google, no own passwords.
  • BFF pattern: the browser never sees a token, only an HttpOnly / Secure / __Host- cookie; CSRF protection.
  • Access by admin approval, one-link invitations; permissions are read from the database on every request.

AI you can steer and afford

  • A pipeline instead of one expensive call: a cheap model (Haiku) triages in batches, a stronger one (Sonnet) scores and writes letters.
  • Structured output with a JSON schema and validation – AI output is never trusted blindly; CVs and ads are untrusted data for the model.
  • Every call is recorded with tokens and price; per-click limits (rolling 24 h) and an emergency brake for the whole app.
  • Own narrow LLM port – the model is switched in configuration, the app does not know the vendor SDK.

Cloud and operations

  • The whole environment as code (Bicep), deployed with one script.
  • One Docker image (Angular + API + worker + headless Chromium) from GitHub Actions; Azure Container Apps + a Container Apps Job for the morning run.
  • Secrets only in Key Vault via managed identity; PostgreSQL, data in the EU; e-mails via Azure Communication Services with SPF and DKIM.

Quality and security

  • 600+ automated tests (533 .NET + 85 Angular) with no network and no AI calls; architecture tests guard the layers.
  • CI on every commit, CodeQL, Dependabot, vulnerable package checks, security headers, rate limiting.
  • A modular monolith with ports and adapters; 21 ADRs public on GitHub.
Backend.NET 10 · ASP.NET Core · EF Core
FrontendAngular 21 · signals · CZ / EN
AIClaude API · Haiku + Sonnet
CloudAzure Container Apps · PostgreSQL · Key Vault

What the agent got wrong and how I found out →

Want to know what AI could do with your system?

Let’s discuss it on a free call.

Book a free call