Resume Engineering

Résumé Engineering

Treat the résumé as a build artifact: facts live in one source-of-truth profile; each résumé is a generated, verified output. Never edit the PDF of record directly, and never let a generated copy become the master.

1. Source of truth

• PROFILE.md — every verifiable fact: education, roles with exact dates, research with real methods and numbers, skills actually owned.
• PROJECTS.md — per-project write-ups plus ready-to-paste bullets.
• Résumés (.tex → .pdf) are outputs. When a fact changes, fix the profile first, then regenerate; a stale source of truth silently spreads into every future variant.
• Keep the original evidence (thesis, posters, manuscripts) archived beside the profile — cite from it, don't duplicate facts out of it.

2. Format that survives parsers

• One page for early-career; conservative layout: single column, a standard serif, black and white, no icons, tables, or text boxes.
• Disable hyphenation (\hyphenpenalty=10000 in LaTeX) — a line-break hyphen splits a keyword and the parser misses it.
• Un-blob joined tokens: PD×LGD×EAD or KYC/CDD parse as one unknown word; space them (KYC / CDD) so each term matches.
• Spell out a high-value acronym once (e.g. "Probability of Default (PD)"), then use the acronym; skip expansions for universal terms. Mirror a target JD's phrasing when it differs.
• Check the compiler's exit code; a failed compile silently leaves a stale PDF that looks current.

3. ATS tuning with an honesty ceiling

• Scan with a keyword tool (exact + acronym-variant matching against a curated bank; coverage % with qualification bands; format checks; keyword-stuffing detection). Coverage target ≈ 90%, not 100%.
• The ceiling: never add a keyword for a skill the candidate does not own. Modern ATS flags stuffing, and every listed term is an interview question waiting to happen. Keep an explicit list of deliberate misses (terms the scan wants but the candidate can't defend) so the gap is a decision, not an oversight.
• Run the scanner in JD mode against each specific posting before applying; fix parse problems (blobs, hyphenation) before considering content changes.

4. Per-track variants — reframe, never fabricate

One profile, N variants, one per target track. A variant differs by:

• Selection — drop the section least relevant to this reader to hold one page.
• Reframing — restate the same real experience in the track's language (e.g. issuer review → due-diligence and escalation language for a compliance track; a scored model → "statistical ML pipeline" for a data track). The facts do not change; the vocabulary does.
• A targeting summary only where the variant needs it (career-changer tracks); it may name the target honestly ("seeking a first role in X") — it may not claim experience that doesn't exist. "Self-studied X" as a qualification line is an amateur signal; a real certification is the only legitimate hard-keyword shortcut, and its entry stays commented out until enrollment actually happens.

5. Verification loop

Every change runs: regenerate → confirm page count → ATS scan (general + JD mode) → read it aloud for naturalness (keyword-soup bullets fail human screens even when they pass parsers). Log the coverage number and what was deliberately left out.