I gave an AI my job hunt. Here's the system that kept it honest.

Yesterday I applied to 12 engineering jobs. I didn't open a single application form.

My AI chief of staff did it — research, verification, form-filling, submission, audit trail. I woke up to 12 submitted applications, 8 dead listings caught before I wasted a click, and 3 forms flagged as broken instead of faked.

This isn't a "10x your job hunt with ChatGPT" post. It's about the verification layer that makes delegating to an AI actually work — because an AI that applies to jobs wrong is worse than no AI at all.

The architecture

The system has four stages, and the important thing is what happens between them:

1. Research — find candidate roles across job boards, company career pages, LinkedIn. Rank by hire probability, not salary.
2. Live verification — open every listing and confirm it's real: no expired banner, an actual Apply control, remote terms that include Nigeria. Yesterday: 8 of 20 listings were dead on arrival. Stale aggregators are the norm, not the exception.
3. Application — fill forms with verified facts only. Resume, GitHub, portfolio attached. Every answer grounded in something checkable.
4. Audit — every submission logged in a tracker with the confirmation (or the honest lack of one), timestamped, with caveats noted.

The honesty constraint

Here's the part that surprised me. Asked about production LLM-API experience, the AI wrote: "No production LLM-API experience." Asked about shipped autonomous agents: "None." Asked about production Go: "No production Go."

It disclosed what I haven't done instead of inventing it. That's not a limitation of the system — it IS the system. A fabricated application that gets you an interview you can't survive is negative value. An honest application to the right role is the whole game.

The rule I gave it: only those you record are true. Every claim in every application had to trace back to a verified fact — my resume, my GitHub, my site. No invented years of experience, no inflated metrics, no phantom credentials.

The failure modes (all real, all yesterday)

• Broken submit handlers. Three application forms (all on the same company's careers site) filled completely, passed validation, then silently refused to submit — the Send button dispatched and nothing happened. The system stopped, flagged them for manual action, and logged them as Drafting. It did not fake a confirmation.
• Dead listings. 8 of 20 from aggregators were expired, closed, or region-locked. Caught at verification, never shown to me as options.
• The honesty tax. Disclosing "no production LLM experience" on an AI Engineer application probably tanks that application. Good. I'd rather lose honestly than win fraudulently and get exposed in a technical screen.

What I'd do differently

The verification stage needs to run before research finishes ranking — I burned research effort on roles that died at verification. And the broken-form problem needs a human-in-the-loop escape hatch earlier: three fully-filled forms sitting in Drafting is a queue, not a result.

The uncomfortable takeaway

The bottleneck was never my ability. It was my bandwidth. Twenty applications in a day is a data-entry problem, and data entry is exactly what machines are for — as long as the verification layer is ruthless.

The engineers who thrive with AI won't be the ones who prompt hardest. They'll be the ones who build verification into the loop.

---

I'm Odinaka Nnamani — backend engineer (TypeScript, Node.js, PostgreSQL, AWS) and founder of Rootmail, email infrastructure that tells you what happened to every email. Open to remote backend/infrastructure roles: heypascal.me