Gael Blair-Mbala

It started from a line in my notes: "Got an 11% match on a job posting I'm totally not qualified for: requesting 11–15 years of experience for an entry-level salary." Rather than keep burning application slots on broken listings, I built a pipeline to scrape, filter, and score postings against my actual resume. Then, mostly to learn how, I rebuilt the final ranking step on top of a real IBM Quantum backend.

Framing the shortlist as a QUBO

Once each job has a match score, choosing which handful to actually apply to is a constrained optimization problem: maximize total match quality, capped at 5 applications total and 5 per company (blasting 15 applications at one place gets you flagged, not hired). That's a natural fit for a Quadratic Unconstrained Binary Optimization problem, solved with QAOA on real IBM Quantum hardware.

Three bugs, three root causes

The pipeline ran silently for a while before it sent a single real email.

1. The circuit never returned. QAOA's cost circuit doesn't auto-decompose into simulatable gates, so evaluation silently fell back to a full matrix exponential of the Hamiltonian. Fine under ~10 qubits, effectively infinite past it. Fix: decompose the circuit first → 0.02s per evaluation.
2. The decode step read past the real variables. That first bug had been hiding a second: QUBO constraints add "slack" qubits, and the naive decode read into them as if they were job variables, crashing with an index error. Fix: decode only the true job-variable bits.
3. Every job scored 0%. Once both were fixed, the run finished fast. And returned nothing. Across 8 qubits, per-category similarity multiplies, so unless a job hits nearly every skill category, the fidelity collapses toward zero ("exponential concentration"). Fix: take the 8th root of raw fidelity.

The manual side still turns up bugs

The whole point of this pipeline is to cut down on manual applications, but the ones that still require a human catch real problems too. Applying through what's almost certainly a Workday-based portal (built around one-at-a-time skill entry that drains a surprising amount of patience), I hit a certification dropdown returning mismatched, clipped results mid-search: a search-relevance bug and a text-overflow bug stacked on top of each other, both visible before I'd finished typing. Same instinct as debugging my own code: spot it, document it, write it up clearly enough that someone else could act on it.