TopX — AI Career Intelligence

An editorial dark-themed resume analyzer that combines real ML, transparent heuristics, and a print-ready career report. Built with Flask, scikit-learn, and a Fraunces × IBM Plex Mono design system.

Live demo: topx-ai-resume-analyzer.onrender.com

---

What it does

Upload a PDF resume. In under 10 seconds, you get:

• Real-ML category prediction — TF-IDF + Logistic Regression trained on 2,483 actual resumes across 24 industry categories (89.5% top-3 accuracy).
• Transparent company matching — 35 companies (TCS, Infosys, Google, Amazon, EY, Hindustan Unilever, DHL...) scored by real skill overlap with their published tech stacks.
• Skill extraction — 411 skills across 14 categories (Languages, Frontend, Cloud, AI/ML, Accounting & Finance, Business Operations, Marketing & Sales, etc.) with section-aware ranking.
• Resume score — 0–100 with a transparent breakdown (Skills / Experience / Projects / Certifications / Education / Summary).
• TF-IDF role matching — semantic similarity between your skills and 47 role profiles.
• Skill-gap roadmap — exactly which skills to learn next for each target role.
• Insights engine — strengths, action items, and profile signals (e.g. "Cloud-Native Skill Set", "Operations / Supply Chain Profile").
• Editorial PDF report — native browser print with a dedicated @media print stylesheet. Vector text, magazine-style cover page, no flicker.
• Resume archive — every analysis saved per user, browseable at /history. Re-uploading the same PDF returns instantly via SHA-256 cache.

---

Tech stack

Layer	Tools
Backend	Flask, Flask-SocketIO, Flask-Limiter, eventlet
ML	scikit-learn (TF-IDF + LogisticRegression, RandomForest), joblib
NLP	Section-aware extraction, alias normalization, ambiguous-skill disambiguation
PDF parsing	pdfminer.six
Auth	werkzeug bcrypt password hashing, session-based
Storage	SQLite (Postgres-ready via DATABASE_PATH env var)
Frontend	Vanilla HTML/CSS/JS, Fraunces (display) + IBM Plex Mono (body)
Print	Native window.print() + dedicated @media print styles

---

Quick start

git clone https://github.com/therahul-yo/TopX-AI-Resume-Analyzer-Job-Recommendation.git
cd TopX-AI-Resume-Analyzer-Job-Recommendation

python3 -m venv .venv
source .venv/bin/activate           # Windows: .venv\Scripts\activate

pip install -r requirements.txt
python app.py

Open http://localhost:5001 in your browser.

---

Project structure

.
├── app.py                          # Flask app, routes, extraction pipeline
├── main.py                         # ML loaders, predictors, scoring, insights
├── ml/
│   ├── train.py                    # Trains the company predictor (RandomForest)
│   └── train_classifier.py         # Trains the resume category classifier (TF-IDF + LogReg)
├── models/
│   ├── company_predictor.joblib    # Pre-trained company model
│   └── category_classifier.joblib  # Pre-trained 24-category model
├── data/
│   └── resumes.csv                 # 2,483 Kaggle resumes (industry-labeled)
├── static/
│   └── style.css                   # Obelisk editorial dark design system
├── templates/
│   ├── index.html                  # Landing page (marquee + hero + features)
│   ├── upload.html                 # Drag-drop upload + loading skeleton
│   ├── result.html                 # Career report (with print styles)
│   └── history.html                # Archive of past analyses
├── Book2.csv                       # Legacy placement dataset (~950 rows)
├── Procfile                        # Render: python app.py
└── requirements.txt

---

How the ML works

1. Resume Category Classifier

Input: raw resume text · Output: top-4 industry categories with confidence

TF-IDF (8000 features, ngram 1-2, sublinear)
   ↓
LogisticRegression (C=2.0, balanced class weights)
   ↓
24 categories: IT, Engineering, Finance, Healthcare, Sales, HR, ...

• Train accuracy: 95.1%
• Test accuracy: 68.8%
• Top-3 accuracy: 89.5%
• Trained on the public Kaggle "Resume Dataset" (2,483 resumes).

2. Company Match Predictor

Replaced a weak RandomForest (31% accuracy on synthetic data) with a transparent heuristic:

• 35 companies, each with curated core skills + nice-to-have skills based on their public job postings
• Score = 0.75 × (core_match / |core|) + 0.25 × (nice_match / |nice|)
• Top-tier companies (Google, Meta, Apple) penalize candidates with <2 yrs experience
• Industry alignment from the category classifier gives a +4% bonus
• Output includes the actual match count: "Matched 7/8 core skills · Top-tier"

3. Skill Extraction

• 411 canonical skills across 14 categories
• Section-aware: skills found in the Skills section get 100% confidence; elsewhere 70%
• Alias normalization: react.js → react, sklearn → scikit-learn, k8s → kubernetes, microsoft excel → ms excel
• Ambiguous skill protection: words like Swift, Go, R, sales, communication must appear in the explicit Skills section, preventing false positives like "Swift" matching the company name "Swift ProSys".

4. Role Match Scoring

• 47 role profiles (Python Developer, Data Scientist, Supply Chain Analyst, Accountant, Tax Associate, ...)
• TF-IDF cosine similarity between extracted skills and role-required skills
• Replaces brittle keyword matching with proper semantic similarity

5. Resume Score Breakdown

Out of 100, fully transparent:

Component	Max	Source
Skills	28	Skill count + density
Experience	22	Years detected from regex
Projects	18	Project section bullets
Certifications	14	Certification text patterns
Education	10	Department codes detected
Summary	8	Has dedicated About/Summary section

---

Design

• Aesthetic: "Obelisk" — editorial magazine on a pitch-black canvas
• Typography: Fraunces (italic display serif) × IBM Plex Mono (body & data)
• Palette: True black #000, cream #F5F0E8 text, single molten orange accent #FF6B35
• Layout: Roman-numeral numbered sections (I, II, III…), magazine masthead, marquee ticker
• Motion: Mask-reveal hero, marquee scroll, IntersectionObserver-driven section reveals, score count-up, animated bars
• PDF: Native window.print() with @media print stylesheet — no flicker, vector text, page-break-aware

---

Production-grade features

• Bcrypt password hashing via werkzeug.security
• Auto-login after registration
• Rate limiting — 12 uploads/hour per user (Flask-Limiter, in-memory)
• PDF hash cache — same resume re-uploaded within 24h returns instantly
• Resume archive — /history route, last 30 analyses per user
• Serialized ML models — joblib bundles ship with the repo for fast cold starts
• Postgres-ready — DATABASE_PATH env var falls back to SQLite
• 5 MB upload limit + extension validation
• Section parsing — Skills / Experience / Education / Projects / Certifications / Summary detected from headers
• Native PDF print — no html2pdf.js, no canvas screenshot, no flicker

---

Re-training the models

If you swap in a new dataset:

# Train the company predictor (uses Book2.csv)
python ml/train.py

# Train the category classifier (uses data/resumes.csv)
python ml/train_classifier.py

# Commit the updated joblib bundles
git add models/*.joblib
git commit -m "Retrain models"

The app loads the serialized bundles at startup. If they're missing, it trains inline as a fallback.

---

Roadmap

Doable on the free tier with no external services:

• Serialized ML models for fast cold starts
• Resume archive with hash caching
• Rate limiting
• Editorial PDF report
• Loading skeleton
• Heuristic company matching with transparent scoring
• Non-tech resume support (commerce, finance, operations)

Needs paid tier or external accounts:

• Postgres for persistent archive (Render free Postgres / Supabase)
• sentence-transformers for semantic skill matching (~500 MB model, needs paid Render)
• spaCy NER for proper resume entity extraction
• Sentry error tracking
• SendGrid email PDF reports
• OG image generation for social shares (Pillow)

---

License

MIT License — free to use, modify, and ship.

---

Made by Rahul · 2025