Application Optimizer

Overview

Applicant tracking systems screen many applications before a person reads them, mostly by matching the CV against the posting’s keywords. Tailoring a CV by hand for every application is slow, and doing it carelessly either misses the keywords or overstates experience. Application Optimizer automates the tailoring while trying to stay honest about fit. It is a single Python script: put the posting in a text file, keep the master CV in LaTeX, run one command, and get a folder with exactly two PDFs.

How the pipeline works

The first step is an analysis call. The model reads the posting and the CV and returns JSON with the role’s seniority, the years of experience it asks for, the required and preferred keywords, a qualification rating from strong match to mismatch, a match score and an optional warning. In strict-fit mode, a stretch or mismatch rating ends the run before any document is written.

Coverage is then measured in code, not by the model: the script checks which of the posting’s keywords already appear in the CV and warns when required coverage is low enough to risk being screened out. The tailoring call receives the missing keywords with an instruction to include them only where the CV supports them. It may reorder bullets by relevance and must keep the LaTeX structure intact, never invent experience and never inflate years.

After tailoring, the script counts how often each required keyword now appears. If any appears too rarely, it re-tailors with just those keywords, for a bounded number of rounds, and it warns when a keyword appears so often that it looks like stuffing.

The cover letter is written as plain text with a word limit and rules against generic praise and exaggeration. For a senior role and a junior candidate, it acknowledges the growth stage instead of pretending otherwise. A filter then strips stock phrases such as “passionate about” or “cutting-edge”, the text is escaped for LaTeX, wrapped in a letter template and compiled. Each application gets its own folder named after the company, and every intermediate file is deleted, so only the two PDFs remain.

What came next

The optimizer’s honesty rules live entirely in its prompts. The script measures keyword counts, but nothing checks whether a tailored bullet still matches what the master CV says. Job Hunt was built on exactly that gap: it requires CV achievements to be copied verbatim from the original, has a second model audit every claim, and stops rather than publishing a draft that fails. The optimizer is where the workflow started, and Job Hunt is where its main weakness was fixed.