Muaz Husaini

← All workCase 02 / 07

Independent build

A job-application copilot that cannot lie about you

Every sentence it writes must cite a fact from my real CV, and a person always presses Submit.

Role
Product, architecture and review
When
2026
Status
Built, pre-launch
Engraved collage: a typewriter, a stack of sealed envelopes, a pinned map, a rubber stamp and a pocket watch.

The problem

Most AI resume tools happily invent experience. A tool that writes about you has to be unable to make things up, and it should never apply anywhere without you.

What I built

A career knowledge base is loaded from my real CV, with an ID on every fact. Tailoring may only use linked facts: each generated bullet cites a fact ID, and a verification pass rejects any claim without a source before a deterministic renderer produces the PDF and DOCX.

Airflow schedules job discovery. Playwright pre-fills applications in copilot mode, which means it stops before the final step and a human always clicks Submit. Supabase runs with row-level security from day one.

I wrote the design, specs and acceptance criteria, and reviewed the code that coding agents implemented against them.

  1. job description
  2. analyserequirements and fit
  3. tailorfrom cited facts only
  4. verifyunsourced claims rejected
  5. render, then a human submitsPDF and DOCX

Where it stands

Knowledge base, tailoring and dashboard are built, with 129 collected tests. The next milestone is the first live discovery run on JobStreet.

  • Python
  • Claude Agent SDK
  • Supabase
  • Airflow
  • Playwright
  • WeasyPrint