ATS match scoring and in-place resume tailoring for any job posting.
Upload your resume PDF and paste a job description. FitPress scores how well the resume matches the job and explains every point. One click then rewrites your lines to match the job and writes them back into your own PDF, in your own layout and fonts. Bullets never gain a number, tool or employer they did not already have. Tailoring also adds missing skills the job asks for to the matching line of your skills section (a language to Languages, a tool to Tools, a concept such as CS Fundamentals to a Coursework or general Skills line, or else to another skills line next to related concepts; never to Languages or Certificates, and never twice); each one is marked so you can untick the ones that are not true. Lines under the same role or project never open with the same verb or gain the same new term twice.
This project replaces the earlier Streamlit "AI Resume Screening and Ranking System".
Scoring (backend/app/scoring)
| Part | Weight | How it is measured |
|---|---|---|
| Skills match | 70% | With a Gemini key, Gemini reads the posting and lists every skill it asks for, including ones inside sentences. Without a key, or if Gemini fails, a built-in list of about 290 skills is used. Either way, names are matched with aliases (JS → JavaScript, K8s → Kubernetes). Required, core and nice-to-have skills are weighted differently. "Django or FastAPI" counts once. A skill shown inside a role earns more than one only in a list. A skill implied by another (GitHub Actions → CI/CD) earns 70%. With Gemini and the embedding model, a practice described in a bullet without being named ("built 40+ internal endpoints" for REST APIs) earns 50%; named tools and languages must appear by name. Soft skills (teamwork, communication) and general ways of working that are not the core of the role (requirements gathering for an engineer) earn credit when shown but are never counted as missing. When a posting names fewer than 5 technical skills, this part counts for proportionally less. |
| Requirement coverage | 10% | Each requirement line in the job is matched to the closest line on the resume. Uses a sentence-embedding model when installed, otherwise skill and word overlap. |
| Keyword match | 2% | Non-skill words the posting repeats (for example "payments"). |
| Experience | 8% | Years computed from job date ranges, merging overlaps, compared with the years the job asks for. |
| Job title | 2% | Closest title on the resume compared with the job title. |
| Education | 1% | Degree level and field compared with the requirement. |
| ATS readability | 7% | 12 checks: selectable text, clean characters, contact details, standard headings, single column, tables, images, reading order, length, dates, measurable results and action verbs. |
Parts the job does not mention (no degree requirement, for example) are left out and the rest are re-weighted. If two or more required technical skills are missing, the score is held at 79, below Strong match. The report also gives a prioritised list of fixes.
Tailoring (backend/app/tailoring)
- Gemini receives each editable line (bullets, summary, skills lines) with the number of characters that fits its space, plus the job's requirements.
- Every rewrite goes through a guard that rejects it if it adds a number, a skill the role or project does not show, or a name that is not on the resume, or if it drops items from a skills list.
- Rewrites that are too long or too short for their lines are sent back once with the exact length to hit.
- Without a Gemini key, or if Gemini fails, an offline mode reorders skills lines so the job's skills come first and names implied terms (for example adds "CI/CD" next to GitHub Actions).
In-place PDF editing (backend/app/pdf)
The original lines are removed and the new text is set at the same positions, size, colour and weight, including bold runs and justified paragraphs. Gemini returns plain text and bold is applied in code to at most four highlights per line: impact numbers first, then skills the job asks for, then other tools and languages, drawn in your resume's own bold font. Skills lines are never bolded. The text is written back at the same point in the PDF's content stream, so ATS parsers still read it in the right order. If the PDF only embeds some letters of its font, the closest font in the same weight is used: the same font if it is installed on the machine, otherwise a bundled one (Latin Modern for LaTeX resumes, Carlito for Calibri, Caladea for Cambria, Lato for Lato) and the app says so. Light, regular and bold are matched separately, so a line set in a light font is not redrawn heavier. For other fonts, install the font or point RESUME_FONT_DIRS at a folder that has it. The text layer is corrected so rewritten lines read back with ordinary spaces and hyphens. After writing, every untouched line is checked to still have the same text in the same place. Any edit that would disturb other text is dropped.
Lines with links, superscripts or mixed colours are left alone.
Requires Python 3.11+ and Node 20+.
From the repository root:
python3 -m venv .venv
source .venv/bin/activate
pip install -r backend/requirements-dev.txt
cp backend/.env.example backend/.env
npm --prefix frontend installAdd your key to backend/.env (free key at Google AI Studio):
GEMINI_API_KEY=your-key
Then, with the virtual environment active:
./scripts/dev.shRun the tests with cd backend && pytest.
Open http://localhost:5173. In development there are two servers: the frontend on port 5173, which is the one you open, and the API on port 8000. The frontend forwards every /api request to the API, so you never need to open 8000 yourself.
To serve everything from one process instead, run ./scripts/serve.sh and open http://localhost:8000. It builds the frontend and FastAPI serves both the app and the API on that one port.
For better requirement matching, install the embedding model as well: pip install -r backend/requirements-ml.txt. It downloads all-MiniLM-L6-v2 on first start and falls back to word matching until it is ready.
| Name | Default | Meaning |
|---|---|---|
GEMINI_API_KEY |
empty | Enables reading the job description with Gemini and rewriting of bullets and summaries. Only the job description is sent for scoring. |
GEMINI_MODEL |
gemini-3.5-flash,gemini-2.5-flash |
Models to try, in order. |
SEMANTIC_BACKEND |
auto |
lexical skips the embedding model. |
MAX_UPLOAD_MB |
8 |
Largest accepted PDF. |
SESSION_TTL_MINUTES |
60 |
How long an uploaded resume is kept in memory. Resumes are never written to disk. |
CACHE_DIR |
backend/.cache |
Where Gemini's reading of each job description is saved, so the same posting always scores the same. |
CORS_ORIGINS |
http://localhost:5173 |
Web addresses allowed to call the API, comma-separated. This is where the frontend runs, not the API's own port. Add your site's address if you host the frontend elsewhere. |
RESUME_FONT_DIRS |
empty | Extra folders to search for fonts that match your resume. |
cd backend
python -m pytestThe fixtures in backend/tests/fixtures are two fictional resumes: one built from a LaTeX template (src/sample_latex.tex) and one exported from Word (src/make_word_fixture.py).
backend/
app/
api/ FastAPI routes, request models, dependencies
core/ settings, date parsing
pdf/ PDF reading, structure detection, layout fitting, in-place editing
scoring/ skills taxonomy, job parsing, scoring engine, ATS checks
tailoring/ Gemini client, no-invention guard, tailoring flow
services/ session store and the analyse → tailor → select workflow
data/ skills.tsv taxonomy
assets/fonts/ substitute fonts and their licences
tests/ unit and integration tests
frontend/
src/
app/ app shell and state
features/ intake, report, proof (page preview and edits), results
components/ shared pieces
hooks/, lib/, styles/
scripts/ dev and serve helpers
docs/ progress notes
| Method | Path | Purpose |
|---|---|---|
| GET | /api/health |
Rewriter in use and embedding status |
| POST | /api/analyze |
Form fields resume (PDF), job_description, optional job_title. Returns a session id and the report. |
| POST | /api/sessions/{id}/tailor |
Proposes and applies edits, returns before and after reports |
| POST | /api/sessions/{id}/selection |
Body {"accepted": [unit ids]} to keep only some edits |
| GET | `/api/sessions/{id}/pages/{original | tailored}/{page}.png` |
| GET | /api/sessions/{id}/download |
Tailored PDF |
- Only text PDFs. Scanned resumes need OCR first.
- Rewritten lines keep their line count, so a rewrite can be at most as long as the space it had.
- Two-column resumes are scored, and flagged, but editing is tuned for single-column layouts.