defensible-cv

English · Tiếng Việt

A profile-driven pipeline that turns one input file into a recruiter-ready, 2-page developer CV PDF — where every number is verified live (publication downloads, citations, GitHub stars/forks) and committed as proof.

---

What this is

You fill in profile.yaml (identity + research sources + experience). A Python crawler verifies the metrics against live public APIs and writes data/research.{json,md}. Then the CV is authored into cv_data.yaml (rendercv) and rendered to a clean 2-page PDF.

profile.yaml  ──►  scripts/research.py  ──►  data/research.json + research.md
                                                      │
                                                      ▼
                                  cv_data.yaml  ──►  rendercv  ──►  PDF (2 pages)

Sample output: assets/Nguyen_Phuong_Anh_Tu_CV.pdf.

Use it with an AI agent (recommended)

This repo is built around an agent skill. Open it in any AI coding agent that reads AGENTS.md / SKILL.md files (Claude, Cursor, …). The agent reads AGENTS.md → agents/skills/defensible-cv/SKILL.md, and stops to ask for profile.yaml if it's missing — it never invents your data.

Then just ask, in plain language:

Build a CV for a new person (no profile yet):

I want a developer CV. Here's my info — name: …, location: …, email: …, GitHub: …, publications (DOIs): …, experience: …, education: …, skills: …. Follow AGENTS.md + the defensible-cv skill: write profile.yaml, refresh the metrics, author cv_data.yaml, and render the PDF.

Rebuild from an existing profile.yaml:

Build the CV: follow AGENTS.md + the defensible-cv skill. Refresh metrics from profile.yaml, then render the PDF.

Refresh the numbers only:

Refresh the CV metrics and re-render the PDF.

Do it manually (no agent)

# 1. Create your profile from the template, then fill it in
cp profile.example.yaml profile.yaml

# 2. Verify metrics (reads profile.yaml's research_sources; --no-scholar = CI-safe)
uv run --with requests --with pyyaml --with pypdf \
  python scripts/research.py --no-scholar      # --refresh bypasses 24h cache

# 3. Render the CV to PDF (no venv needed)
uv run --with "rendercv[full]>=2.8" rendercv render cv_data.yaml

scripts/research.py reads only profile.yaml's research_sources. If profile.yaml is missing it stops with a clear message instead of inventing data. The committed profile.yaml / cv_data.yaml are a complete worked example.

How it stays defensible

Every figure on the CV is verified live, then committed to data/research.json so it survives a recruiter's spot-check:

Source	What it returns
OJS (e.g. jte.edu.vn)	Per-article download total from the inline pkpUsageStats payload.
Semantic Scholar Graph API	Paper metadata + paginated citing-paper list (venue, year, authors, DOI).
OpenCitations COCI/Meta	Independent citation count + citing DOIs (cross-checks Semantic Scholar).
Crossref Event Data	Twitter / Wikipedia / news mentions (best-effort).
Google Scholar (scholarly)	Optional, soft-fails on CAPTCHA. --no-scholar skips it.
GitHub Public API	Per-repo stars/forks, languages, totals across non-fork repos.
Citing-paper PDFs (Unpaywall → arXiv → pypdf)	Downloads each open-access citing paper, finds the reference label that points back to the work, and extracts the verbatim sentence + page where the author invokes it.

Concrete citation contexts

For every open-access citing paper, data/research.md records the reference label, the reference-list page, and the in-body sentence(s) that cite the work. Example from the committed run for 10.54644/jte.2024.1514:

[ACM] Integrating Expert Knowledge With Automated Knowledge Component Extraction for Student Modeling — ACM UMAP 2025

• Our paper is reference [19] (their p.6).
• p.2 — "…ASTs have been widely used in automated code analysis efforts [19]. For example, Rivers used ASTs to identify each syntax structure in a student's submission…"

Paywalled papers (most IEEE, parts of ACM, MDPI behind Akamai) are kept with status not_oa / fetch_failed and clearly flagged — nothing is fabricated.

Files

profile.yaml                 INPUT: identity + research sources + narrative
profile.example.yaml         documented template — copy to profile.yaml
cv_data.yaml                 rendercv source (authored from profile + research)
scripts/research.py          CLI crawler, single file, urllib + pypdf
data/research.json           machine-readable verified snapshot (regenerated)
data/research.md             recruiter-friendly summary (regenerated)
AGENTS.md                    entry point for AI agents
agents/skills/defensible-cv/SKILL.md   the agent playbook
assets/cv-preview.png        the preview banner above
assets/Nguyen_Phuong_Anh_Tu_CV.pdf     sample rendered CV
requirements.txt             pinned deps

License

MIT.