ResumeScorer

Local, private resume scoring + optimization lab. Wraps interviewstreet/hiring-agent (HackerRank's resume-to-score pipeline) and runs it fully offline on local Ollama — plus a Claude Code skill + agent that turn raw scores into honest, prioritized resume edits.

Primary use: reverse-engineer what automated screeners reward and raise Richard's resume score. Also: score/benchmark candidate resumes, and study how the PDF→JSON→GitHub→score pipeline works.

---

What's where

ResumeScorer/                         <- this project (docs + scripts + results)
├── README.md                         <- you are here
├── CLAUDE.md                         <- agent harness for working in this project
├── AGENTS.md                         <- how to run agent sessions against this system
├── docs/
│   ├── user-manual.html              <- full visual user manual (open in a browser)
│   ├── rubric-explained.md           <- how the 120-pt score works + leverage
│   ├── next-steps.md                 <- grounded next actions from the benchmark
│   └── handoff/                      <- append-only session handoff tree
│       └── HANDOFF.md
├── skills/score-resume/SKILL.md      <- source-of-truth copy of the Claude skill
├── agents/resume-optimizer.md        <- source-of-truth copy of the Claude agent
├── scripts/
│   ├── score_resume.ps1              <- score ONE resume
│   └── batch_score.ps1               <- score the whole corpus
└── results/                          <- score reports (.txt), logs, summary, CSV

C:\Users\a2021\dev\hiring-agent       <- the actual pipeline code + venv (OUTSIDE OneDrive on purpose)

The Python repo lives in C:\Users\a2021\dev\hiring-agent, not in OneDrive — a venv + git + model cache must not sync. This project is the docs/control/results hub that wraps it.

Setup state (already done — 2026-06-25)

Component	State
Repo clone	C:\Users\a2021\dev\hiring-agent
Python	3.11 venv at …\hiring-agent\.venv (repo pins 3.11.13)
Deps	requirements.txt installed
Backend	Ollama, CPU mode (gemma3:4b). GPU disabled — see below
Config	…\hiring-agent\.env → LLM_PROVIDER=ollama, DEFAULT_MODEL=gemma3:4b
Skill	score-resume installed at ~/.claude/skills/
Agent	resume-optimizer installed at ~/.claude/agents/

⚠️ GPU is disabled (CPU-only) — and why

The RTX 4060 is present, but the NVIDIA driver is 546.01 (late-2023), too old for Ollama 0.30.9's bundled CUDA runtime. The GPU path crashes (CUDA error: device kernel image is invalid). We forced CPU via the user env var OLLAMA_LLM_LIBRARY=cpu_avx2 (Ollama-scoped — does not hide the GPU from your other ML tools). CPU works fine on 32 GB RAM; it's just slower. To re-enable GPU later: update the NVIDIA driver to ≥ 550, then remove the OLLAMA_LLM_LIBRARY user env var and restart Ollama. See docs/next-steps.md.

Quick start

Score one resume (from anywhere):

& "C:\Users\a2021\dev\hiring-agent\score_resume.ps1" -Pdf "C:\path\to\resume.pdf"

Score the whole corpus (Richard's 4 + ScoutLane samples):

& "…\ResumeScorer\scripts\batch_score.ps1"

Or, in Claude Code, just say "score my resume" / "optimize my resume" — the score-resume skill handles the rest.

The 30-second mental model

PDF → (LLM extracts sections to JSON) → (GitHub repos pulled if a GitHub URL is on the resume) → (LLM scores against a 120-pt rubric) → printed report + CSV row. 65% of the score is open-source + self-projects, both GitHub-verified — so the highest-leverage fix for most engineering resumes is making real work legible (GitHub link, live demos, quantified impact). Full detail: docs/rubric-explained.md.