🚧 Work in progress. This is an active build by Gradian — usable today, still evolving. Expect rough edges.
A local, agentic CV ↔ job-offer analyzer. It reads a CV and a job offer and tells you — explainably — how well they fit, then rewrites the CV to fit better. It works from both sides of the table: the Applier (job-seeker) and the Recruiter (hiring).
It runs entirely on your machine and on your own AI account. Nothing is stored; nothing leaves your computer except the calls you make. It is not a website you log into — it's an app you run.
(Full walkthrough: docs/how-it-works.pdf.)
- Explainable compatibility score — ATS keyword coverage + recruiter gating (languages, years) + the AI's semantic read, each shown separately so you see why, not just a black-box %.
- Job Finder across free job boards (Arbeitnow, Remotive; add Adzuna/Jooble keys for more), or paste a specific offer link.
- Regenerate your CV for a specific offer with a 1–100 aggressiveness slider, refined by an internal Tailor ↔ Critic loop (keeps rewriting until it passes the quality gate), plus a transparency truth-ledger of anything newly claimed — and a one-click ATS-safe PDF (previewed inline).
- Screen & rank a batch of candidate CVs against one role.
- AI-writing signals — heuristics + a reading model flag likely AI-written CVs as signals, never proof.
- Source verification — checks that a candidate's links and GitHub repositories are real.
- Legal sourcing — GitHub user search + a Google/LinkedIn X-ray string you run in your own browser (no scraping).
Every run streams a live agent console — you watch the Extractor, Analyst, Tailor, Critic (and, for recruiters, Examiner, Verifier, Sourcer) light up as they work, with a progress bar. No more staring at a frozen button.
- An AI backend — pick one:
- Claude Code (recommended, no key): install and sign in — https://claude.com/claude-code (check with
claude --version). - Anthropic API key (faster per call): put
ANTHROPIC_API_KEY=…in.env(get one at https://console.anthropic.com/).
- Claude Code (recommended, no key): install and sign in — https://claude.com/claude-code (check with
- Python 3.11+
- Google Chrome or Microsoft Edge — used to render the CV to PDF.
- Download or clone this repo.
- Double-click
start.bat— the first run creates a virtual environment and installs dependencies automatically. - Your browser opens at http://127.0.0.1:8765.
Paste (or upload a PDF of) your CV and a job offer, then Analyze. Move the slider and Regenerate to get a tailored CV you can download as a PDF. Switch to the Recruiter tab to screen candidates.
On other platforms, run it by hand:
python -m venv .venv→ activate →pip install -r requirements.txt→PYTHONPATH=src python -m uvicorn gradianmatch.server:app --port 8765.
Copy .env.example to .env:
| Key | For |
|---|---|
ANTHROPIC_API_KEY |
Use the API backend instead of Claude Code. |
GM_BACKEND |
Force auto (default) / cli / api. |
GM_MODEL |
Override the model (default claude-opus-4-8 on the API backend; e.g. claude-sonnet-5 for faster/cheaper). |
ADZUNA_APP_ID / ADZUNA_APP_KEY / JOOBLE_KEY |
Wider Job Finder coverage. |
GITHUB_TOKEN |
Higher GitHub rate limits for source verification. |
The analyze / regenerate core needs none of these — only an AI backend.
Everything runs locally and in-memory per session — nothing is persisted, temporary files are deleted. The server binds to 127.0.0.1 (your machine only) and rejects cross-origin requests; do not expose it to a public network. The only text that leaves your machine goes to your own AI account and any job boards you enable.
The tailoring slider can emphasize your real experience aggressively, but the truth-ledger flags anything it adds beyond your source CV — you are responsible for the accuracy of what you send to employers. In recruiter mode you process third parties' data: the legal basis is yours, and the tool holds nothing. The AI-writing flags are signals, not proof — never auto-reject anyone on them.
- Deterministic core (extraction, skill taxonomy, scoring, source verification, job-board adapters) with no AI, plus an agentic layer (Analyst / Tailor / Critic / Examiner / Sourcer as prompt modules) orchestrated by a FastAPI backend that serves a buildless vanilla-JS UI.
- All AI goes through one door —
ai_client→ Claude Code CLI or the Anthropic SDK. - Run the tests:
pip install -r requirements-dev.txtthenpython -m pytest.
This tool applies the same production pattern as the rest of Gradian — narrow specialist agents, a producer ↔ critic loop with a numeric quality gate, and a human making the final call:
| Repo | What it shows |
|---|---|
| gradian-sistema | The factory: an interactive map of the 28-agent office, with measured usage numbers. The system. |
| gradian-caso-waterfront | A finished piece from that factory: a real pub's website, live in production, explained from the paper menus to deploy. The result. |
| gradian-match (you are here) | The same pattern transferred to another domain: CV ↔ job-offer analysis. The pattern, transferred. |
| gueridon | Menu-engineering science + a live menu analyzer: tested deterministic core, agentic shell with per-stage fallback. The product. |
Source-available, not open source. © Gradian. All rights reserved — see LICENSE. You may read and run it locally for personal, non-commercial use; redistribution, commercial use, and derivative distribution require written permission. Questions: alex@gradiangrowth.com.