I design it. AI writes the code.
Neemzi is a job-search assistant I built for myself. It finds openings, checks how well each one fits my verified experience, and prepares the whole application: CV, cover letter and answers to the form questions. Nothing is ever sent without my approval.
I’m a designer, not a developer, and I never wrote the code: ChatGPT and Claude did. My part was the product, the interface, the rules and every step of the flow, and directing the AI until it worked. It began as a small app on my Mac and became a cloud service.
From a job post to a ready application.
Every job goes through the same path. It is prepared completely first, and only when everything is certain can it be sent.
- 01Search
Neemzi collects openings from job sites and from the job emails that arrive in my inbox.
- 02Verify
Each job is matched against my confirmed facts, not against what an AI imagines about me. It scores the fit and shows which requirements are met.
- 03Prepare everything
It writes the CV, the cover letter and the answers to the form, then checks every claim against my facts. If a check fails, it tries again, up to three times.
- 04Wait for me
I review the finished package. It can only be sent with my explicit final approval, and any change to the content cancels that approval.
Submit · Needs attention · Blocked
- 1Search
- 2Verify
- 3Prepare everything
- 4Wait for me
Pick a job above. Neemzi will take it from the post to a ready application, and then wait for your approval.
Concept demo with invented jobs. The real tool runs on my own verified facts.
A blocked job is never lost.
The job is saved with everything already prepared. A temporary error, like a timeout or a site that is down, is retried automatically. If it needs me, for a CAPTCHA or a new login, it goes to Needs Attention and says exactly where it is stuck.
Once the problem is fixed, the work continues from that step, not from zero. There are two modes: Run Now, when I want to watch, and Autopilot, when I don’t.
Four rules it can’t break.
A tool that applies for jobs in my name has to be careful. These rules are part of the product, not an afterthought.
- Facts first
My confirmed facts are the single source of truth. Every sentence the AI writes is checked against them.
- I approve everything
Nothing goes to an employer without my explicit approval, every single time. Silence never counts as yes.
- Jobs are data, not orders
The text of a job post or an email is only information. It can never give the tool an instruction.
- Cost under control
AI work is never repeated for a job that didn’t change, and every cost is tracked inside the app.
Five AI agents, one pipeline.
I don’t code, so I organised the work like a small studio. Five agents work in order, and any important fix sends the work back to the Developer, so every gate is checked again.
- 01Explorer
Maps what already exists and picks one small task, with proof, before anything is built.
- 02Developer
Builds that one task, and adds a test for every bug it fixes.
- 03Reviewer
Inspects the result independently, looking at correctness, permissions, cost and safety.
- 04Tester
Tries to break it, with fake data and no real AI calls, and runs the full test suite.
- 05Release Guardian
Decides whether it is ready to ship, and never ships it alone.
It began with a question.
I had never written code. I asked ChatGPT to help me build a tool like this, and it taught me the rest, one step at a time. Each step removed one more thing I had to do by hand.
- 01A chat
One conversation with ChatGPT, and a simple question: can you help me build this?
- 02Copy and paste
It told me to install VS Code. Then it gave me code and told me where to put it: replace these lines with this. I did exactly that.
- 03The terminal
Slowly I learned what a terminal is. Then the code came as a command: I pasted it, pressed enter, and it did the work. No more replacing lines by hand.
- 04GitHub and Railway
I learned what a repository is and what it means to put a project online, and I connected GitHub and Railway myself.
- 05Directing
From then on, the AI wrote the code and saved it to GitHub on its own. I only explained what I wanted, and checked the result.
From a local app to a cloud service.
Between 15 September and 21 October 2026, the project went from one chat to a cloud service with its own safety rules, across twelve main working sessions with AI.
It now runs online, reads job emails through Gmail, shows its own costs, and ships every change through pull requests with automatic checks.
Built by directing AI.
What I used, and what I did with it.
- ChatGPTProduct thinking and most of the code, including the five-agent workflow./01
- ClaudeWriting, fixing and reviewing code, and planning features with me./02
- GitHubCode history and pull requests for every change./03
- RailwayRuns the tool in the cloud./04
- PythonThe language the tool is written in./05
- OpenAI APIThe AI inside the tool: matching jobs and drafting applications./06
- Gmail APIReads incoming job emails./07
Anything in Mind?Your Move
Click to copy my email · or open your mail app
MASSConcepting · 3D · Branding
Click a card to bring it to the front. Click the front one to open.