Live Demo: calibr-two.vercel.app · API Docs: calibr-backend-8vq8.onrender.com/docs
Calibr is a full-stack AI-powered resume screening platform built for recruiters and candidates. It scores resumes using a three-signal hybrid engine — skill graph matching, semantic similarity, and experience relevance — and uses an LLM only for explanation, never for scoring decisions.
Most AI resume screeners let the LLM decide the score. Calibr does not.
Your code computes the score. The LLM only explains it.
Resume + JD
│
├── Signal 1: Skill Graph Match (weight: 0.45)
│ └── Exact match + synonym taxonomy (200+ synonyms)
│
├── Signal 2: Semantic Similarity (weight: 0.35)
│ └── Sentence-Transformers (all-MiniLM-L6-v2)
│
└── Signal 3: Experience Relevance (weight: 0.20)
└── Recency decay + duration weighting
│
▼
Weighted Composite Score (0–100)
│
▼
Groq LLM (llama-3.1-8b-instant)
└── Explains the pre-computed score
Never changes it
This makes the scoring auditable, consistent, and explainable — the same resume against the same JD always gets the same score.
| Metric | Value |
|---|---|
| 20 resumes scored | 0.46 seconds |
| Per resume | 0.02 seconds |
| Weights validated | 5/5 gold-set agreement |
| Skill taxonomy | 200+ synonyms across 15 categories |
- JD Upload — paste job description, auto-parses required skills and experience years
- Single Resume Scoring — upload PDF or paste text, full three-signal breakdown
- Batch Screening — upload 100+ PDFs, Celery processes asynchronously with live progress bar
- Rankings — all candidates sorted by composite score with AI one-line summaries
- Score Breakdown — per-signal analysis showing exactly why each candidate ranked where they did
- AI Interview Questions — 3 technical + 2 gap + 2 behavioral questions tailored to each candidate's profile
- Skill Gap Analysis — per-missing-skill learning resource recommendations
- Recruiter Notes & Shortlisting — shortlist/pending/reject status + private notes per candidate
- Analytics Dashboard — total candidates, avg score, hiring funnel, skill distribution, top 5 candidates
- ATS Score Check — upload PDF resume, full extraction (name, email, phone, skills, experience, projects, certifications)
- JD Dropdown — select from all available job descriptions
- Score + Breakdown — composite score, signal bars, matched/missing skills, personalized feedback
- Skill Gap Resources — Groq suggests learning resources for each missing skill
- Interview Prep — questions you're likely to be asked based on your profile
- ATS Resume Optimizer — bullet rewrites, missing keywords, priority action plan
- Score History — all previous attempts with stats and trend
| Layer | Technology |
|---|---|
| API Framework | FastAPI |
| Database | PostgreSQL (Neon serverless) |
| ORM | SQLAlchemy |
| Task Queue | Celery 5.3.6 |
| Message Broker | Redis (Upstash) |
| NLP | spaCy (en_core_web_sm) |
| Similarity | Sentence-Transformers (all-MiniLM-L6-v2) |
| LLM | Groq API (llama-3.1-8b-instant) |
| PDF Extraction | pdfplumber + PyPDF2 |
| Auth | JWT (python-jose) + bcrypt |
| Rate Limiting | slowapi |
| Layer | Technology |
|---|---|
| Framework | React 19 + Vite |
| Routing | React Router v6 |
| HTTP Client | Axios |
| Styling | Inline CSS (sepia newspaper theme) |
| Service | Provider |
|---|---|
| Backend | Render (Docker) |
| Frontend | Vercel |
| Database | Neon PostgreSQL |
| Redis | Upstash |
| Container | Docker |
| Method | Endpoint | Description |
|---|---|---|
| POST | /auth/signup |
Create recruiter or candidate account |
| POST | /auth/login |
Login, returns JWT access + refresh token |
| GET | /auth/me |
Get current user info |
| POST | /jd |
Upload job description (auto-parses skills) |
| GET | /jd |
List recruiter's JDs |
| GET | /jd/public/all |
All JDs (for candidate dropdown) |
| POST | /score |
Score one resume against a JD |
| POST | /upload/resume |
Extract text + structured data from PDF |
| POST | /batch/score |
Submit batch of resumes (Celery async) |
| GET | /batch/status/{job_id} |
Poll batch job progress |
| GET | /rankings/{jd_id} |
Ranked candidates for a JD |
| GET | /candidates/{id}/breakdown |
Full signal breakdown for one candidate |
| PATCH | /candidates/{id}/status |
Shortlist / reject / pending |
| PATCH | /candidates/{id}/notes |
Add recruiter notes |
| GET | /interview/{candidate_id} |
AI-generated interview questions |
| GET | /skill-gap/{candidate_id} |
Skill gap with learning resources |
| GET | /improve/{candidate_id} |
ATS resume improvement suggestions |
| GET | /analytics |
Recruiter analytics dashboard data |
| GET | /history |
Score history |
Composite Score = (Skill Score × 0.45) + (Semantic Score × 0.35) + (Experience Score × 0.20)
Signal 1 — Skill Match (45%) Compares extracted resume skills against JD required skills using a hand-curated synonym taxonomy (e.g. "ML" = "Machine Learning" = "Artificial Intelligence"). Returns percentage of JD skills matched.
Signal 2 — Semantic Similarity (35%) Neural semantic embeddings using Sentence-Transformers (all-MiniLM-L6-v2). Encodes resume and JD into dense vector representations and computes cosine similarity to capture meaning beyond keyword overlap.
Signal 3 — Experience Relevance (20%) Compares total experience months against JD required years. Applies recency decay — recent experience weighted higher than older experience.
Weights were validated against a 5-candidate gold set with 5/5 agreement before deployment.
- JWT access tokens (30 min expiry) + refresh tokens (7 days)
- bcrypt password hashing
- IDOR prevention — every endpoint verifies
owner_idmatches JWT user - Rate limiting: 10 req/min on
/scoreand/auth/login - Secrets in environment variables, never hardcoded
pip-auditclean — zero known vulnerabilities in dependencies
- Python 3.11+
- Node.js 18+
- Redis (Docker:
docker run -d -p 6379:6379 redis:alpine)
# Clone
git clone https://github.com/GYatharth/Calibr.git
cd Calibr
# Virtual environment
python -m venv venv
venv\Scripts\activate # Windows
source venv/bin/activate # Mac/Linux
# Install dependencies
pip install -r requirements.txt
python -m spacy download en_core_web_sm
# Environment variables
cp .env.example .env
# Fill in: DATABASE_URL, GROQ_API_KEY, SECRET_KEY, REDIS_URL
# Run API
uvicorn main:app --reload
# Run Celery worker (separate terminal)
celery -A app.celery_app worker --loglevel=info --pool=solocd frontend
npm install
npm run devOpen http://localhost:5173
| Variable | Description |
|---|---|
DATABASE_URL |
Neon PostgreSQL connection string |
GROQ_API_KEY |
Groq API key for LLM explanations |
SECRET_KEY |
JWT signing secret |
REDIS_URL |
Redis connection URL (Upstash in production) |
Calibr/
├── app/
│ ├── api/ # FastAPI routers
│ │ ├── auth_router.py # Signup, login, JWT
│ │ ├── jd_router.py # JD upload + listing
│ │ ├── scoring_router.py # Single resume scoring
│ │ ├── batch_router.py # Batch upload + status
│ │ ├── rankings_router.py # Candidate rankings
│ │ ├── candidate_router.py # Score breakdown
│ │ ├── interview_router.py # Interview questions
│ │ ├── skill_gap_router.py # Skill gap analysis
│ │ ├── resume_improve_router.py # ATS optimizer
│ │ ├── analytics_router.py # Recruiter analytics
│ │ ├── history_router.py # Score history
│ │ ├── upload_router.py # PDF extraction
│ │ └── notes_router.py # Notes + shortlisting
│ ├── db/
│ │ ├── models.py # SQLAlchemy models
│ │ └── database.py # DB connection
│ ├── scoring/
│ │ ├── skill_match.py # Signal 1
│ │ ├── semantic_similarity.py # Signal 2
│ │ ├── experience_score.py # Signal 3
│ │ ├── composite_score.py # Weighted combination
│ │ └── explainer.py # LLM explanation + AI features
│ ├── parsing/
│ │ ├── extract_pdf.py # PDF text extraction
│ │ ├── parse_jd.py # JD skill parsing
│ │ └── parse_resume.py # Resume structured parsing
│ ├── celery_app.py # Celery configuration
│ └── tasks.py # Async batch scoring task
├── frontend/
│ └── src/
│ ├── pages/ # React pages
│ │ ├── Login.jsx
│ │ ├── Signup.jsx
│ │ ├── Dashboard.jsx
│ │ ├── Score.jsx
│ │ ├── Batch.jsx
│ │ ├── Rankings.jsx
│ │ ├── Breakdown.jsx
│ │ ├── Analytics.jsx
│ │ ├── CandidateDashboard.jsx
│ │ └── History.jsx
│ ├── components/
│ │ └── Layout.jsx
│ └── api/
│ └── client.js # Axios + JWT interceptors
├── Dockerfile
├── requirements.txt
├── main.py
└── config.json # Scoring weights
Python 3.11 · FastAPI · PostgreSQL · Redis · Celery · React 19 · Vite · Docker · Render · Vercel · Groq API · spaCy · Sentence-Transformers · pdfplumber
Built by Yatharth Gupta — BTech CSE at SRMIST GitHub: GYatharth · Project: GYatharth/Calibr