CareerLens — AI-Powered Career & Burnout Predictor

Live Demo: career-lens-ml.vercel.app

CareerLens is a full-stack machine learning application that predicts three critical student career outcomes — burnout risk, career readiness, and likelihood of seeking career counseling — from a structured set of academic, behavioural, and AI-usage inputs. The system pairs a rigorously built multi-target ML pipeline with a polished React interface and a production FastAPI backend.

---

Table of Contents

• Overview
• Live Demo
• Features
• Architecture
• Machine Learning Pipeline
  • Dataset
  • Data Preprocessing
  • Feature Engineering
  • Feature Selection
  • Model Training & Hyperparameter Tuning
  • Model Serialisation
• Tech Stack
• Project Structure
• Getting Started
  • Prerequisites
  • Backend Setup
  • Frontend Setup
  • Environment Variables
• API Reference
• License

---

Overview

The project is motivated by a real problem: as AI tools become embedded in academic life, students face compounding pressures — dependency on AI, fear of job displacement, placement anxiety, and burnout. CareerLens quantifies these pressures into actionable predictions that can guide career counseling interventions.

Three separate supervised learning models are trained against the AI Dependency & Career Anxiety in Students dataset (15,000 records, 30 features):

Target	Task	Model
burnout_score	Regression	Linear Regression
overall_career_readiness_score	Regression	Tuned XGBoost Regressor
seeks_career_counseling	Binary Classification	Tuned XGBoost Classifier

Each XGBoost model is optimised independently via Optuna Bayesian hyperparameter search, and all three are served through a single FastAPI /predict endpoint.

---

Features

• Multi-target prediction — three independent ML models served behind one API call
• Leakage-free preprocessing — train/test split happens before any imputation or encoding; all statistics are computed on training data only
• Bayesian hyperparameter optimisation — Optuna TPE sampler with MedianPruner across 75–100 trials per model
• Information Value feature selection — Weight of Evidence / IV filtering applied to the classification target to surface only discriminative features
• VIF-based multicollinearity removal — Variance Inflation Factor analysis eliminates redundant predictors before model training
• Logic-driven imputation — categorical missingness in primary_ai_tools_used resolved through structural inference from daily_ai_tool_usage_hrs, not arbitrary fill
• ROC / AUC evaluation — XGBoost classifier evaluated with full ROC curve and AUC reporting
• Class imbalance handling — scale_pos_weight parameter included in XGBoost classifier search to compensate for class skew
• React multi-step form — four-section guided input (Personal → Academic → AI & Career → Lifestyle) with client-side validation
• Fully responsive — mobile-optimised layout with touch-friendly slider and radio components

---

Architecture

┌─────────────────────────────────────┐
│         React Frontend (Vite)       │
│  Multi-step form → POST /predict    │
│  Deployed: Vercel                   │
└────────────────┬────────────────────┘
                 │ HTTPS (CORS-restricted)
┌────────────────▼────────────────────┐
│         FastAPI Backend             │
│  /predict  →  prediction_helper.py  │
│  /model_status                      │
│  Deployed: Render                   │
└────────────────┬────────────────────┘
                 │
┌────────────────▼────────────────────┐
│     artifacts/models.pkl (joblib)   │
│  ├── MinMaxScaler                   │
│  ├── Burnout Regressor              │
│  ├── Career Readiness Regressor     │
│  └── Counseling Classifier          │
└─────────────────────────────────────┘

The frontend is a Vite + React deployed on Vercel. The backend is a FastAPI service that loads a single serialised joblib artifact containing the scaler and all three models.

---

Machine Learning Pipeline

The full pipeline is documented in backend/Career_Predictor_Documented.ipynb.

Dataset

• Source: AI Dependency & Career Anxiety in Students
• Size: 15,000 records × 30 features
• Split: 70% train (10,500 rows) / 30% test (4,500 rows)

The split is performed before any EDA, imputation, or transformation. This is a deliberate architectural decision to prevent data leakage — no test-set information contaminates the statistics derived during preprocessing.

Data Preprocessing

Categorical imputation — primary_ai_tools_used: Missingness in this column is structurally non-random. Rather than collapsing all missing values into a single category, a two-branch rule is applied:

• daily_ai_tool_usage_hrs == 0 → fill with 'None' (student does not use AI tools)
• daily_ai_tool_usage_hrs > 0 → fill with 'Unknown' (student uses AI but the tool is unrecorded)

Numeric imputation: Distribution shapes of sleep_hours, social_media_hrs_per_day, and self_learning_hours_per_week were inspected via KDE plots before selecting a strategy. All three exhibited skew, so median imputation was applied (medians computed on training data only, then applied to test). seeks_career_counseling was imputed with the training-set mode.

Feature Engineering

Column	Encoding	Rationale
college_tier	Ordinal (Tier 3→0, Tier 2→1, Tier 1→2)	Natural rank ordering
uses_ai_for_assignments	Ordinal (Never→0 … Always→4)	Ordered frequency scale
urban_or_rural	Binary (Rural→0, Urban→1)	Binary nominal
gender, degree_type, stream, primary_ai_tools_used	One-hot (drop_first=True)	Nominal, no ordering

df_train2.align(df_test2, join='left') ensures test-set dummy columns missing from training are added and zero-filled, preventing shape mismatch during inference.

Feature scaling is performed with MinMaxScaler fitted exclusively on the training set.

Feature Selection

Multicollinearity check (VIF): Variance Inflation Factor analysis is run on the encoded feature matrix. Features with VIF above the threshold are iteratively removed to reduce redundancy and improve model stability.

Weight of Evidence / Information Value (WoE/IV): IV-based selection is applied specifically to the classification target (seeks_career_counseling). Only features with IV above the minimum threshold are retained in X_train2 / X_test2, the feature matrix used for classifier training. Regression targets use the full feature set (X_train1).

Model Training & Hyperparameter Tuning

Each target is trained with a baseline sweep (Linear/Logistic Regression, Random Forest, XGBoost) followed by Optuna-driven hyperparameter search on the best-performing model family.

Optuna configuration (all studies):

• Sampler: TPESampler(seed=42) — Tree-structured Parzen Estimator Bayesian search
• Pruner: MedianPruner — terminates unpromising trials early
• CV: 3-fold (regression, R² metric) / 5-fold (classification, macro F1 metric)

Burnout Score (Regression): Out of all models LinearRegression showed the best R² score and tells that burnout has a complete linear relationship to other columns and hence LinrearRegression was chosen.

Career Readiness Score (Regression): Search covers ElasticNet (L1 + L2 penalty mix via saga solver, alpha, l1_ratio) and XGBRegressor (n_estimators, learning_rate, max_depth, min_child_weight, gamma, subsample, colsample_bytree/bylevel/bynode, reg_alpha, reg_lambda, max_delta_step, grow_policy). Metric: R².

Career Counseling Classifier (Binary Classification):

• Logistic Regression search: L1/L2/ElasticNet penalties via saga solver, C, class_weight (None / balanced). 75 trials. Metric: macro F1.
• XGBoost search: Same tree search space, plus scale_pos_weight searched in range [1, neg/pos × 2] to handle class imbalance. 100 trials. Metric: macro F1.
• ROC/AUC: Final classifier evaluated with predict_proba soft scores plotted as a full ROC curve.

Model Serialisation

All inference components are bundled into a single artifacts/models.pkl via joblib:

{
    'scaler': {'scaler': <MinMaxScaler>, 'cols': [...]},
    'model_burnout_score': {'model': <best_regressor>, 'cols': [...]},
    'model_seeks_career_counseling': {'model': <XGBClassifier>, 'cols': [...]},
    'model_overall_career_readiness_score': {'model': <XGBRegressor>, 'cols': [...]}
}

Column lists are stored alongside each model so that prediction_helper.py can correctly reorder and subset incoming feature DataFrames before prediction, preventing silent shape mismatches at inference time.

---

Tech Stack

ML / Data Science

• Python 3.10
• pandas, NumPy
• scikit-learn (preprocessing, linear models, ensemble models, metrics)
• XGBoost
• Optuna (Bayesian hyperparameter optimisation)
• statsmodels (VIF)
• seaborn, matplotlib (EDA visualisation)
• joblib (model serialisation)

Backend

• FastAPI 0.136
• Pydantic v2 (request/response validation)
• Uvicorn (ASGI server)

Frontend

• React 18 (hooks only — no class components)
• Vite
• Vanilla CSS-in-JS (zero external UI library dependency)

Infrastructure

• Frontend: Vercel
• Backend: Render / Railway (Python 3.10.13 runtime)

---

Project Structure

soham-lodh-career-helper/
├── backend/
│   ├── Career_Predictor_Documented.ipynb   # Full ML pipeline with inline documentation
│   ├── main.py                             # FastAPI app, routes, Pydantic schemas
│   ├── prediction_helper.py                # Inference logic: encoding, scaling, predict
│   ├── requirements.txt                    # Pinned Python dependencies
│   ├── runtime.txt                         # Python version for deployment
│   └── artifacts/
│       └── models.pkl                      # Serialised scaler + 3 models (joblib)
├── frontend/
│   ├── index.html
│   ├── vite.config.js
│   ├── package.json
│   ├── eslint.config.js
│   └── src/
│       ├── main.jsx                        # React entry point
│       ├── App.jsx                         # Full application: form, steps, results
│       └── index.css
├── LICENSE                                 # MIT
└── README.md

---

Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• npm or yarn

Backend Setup

# Clone the repository
git clone https://github.com/soham-lodh/career-helper.git
cd career-helper/backend

# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Start the development server
uvicorn main:app --reload --port 8000

The API will be available at http://localhost:8000. Interactive docs at http://localhost:8000/docs.

Note: artifacts/models.pkl must be present before starting the server. Either use the pre-trained artifact from the repository or re-run Career_Predictor_Documented.ipynb to regenerate it.

Frontend Setup

cd ../frontend

# Install dependencies
npm install

# Start the development server
npm run dev

The frontend will be available at http://localhost:5173.

Environment Variables

Create a .env file in the frontend/ directory:

VITE_API_BASE_URL=http://localhost:8000

For production, set VITE_API_BASE_URL to your deployed backend URL in your Vercel project environment variables.

---

API Reference

GET /model_status

Returns the current load status of the ML models.

Response

{
  "model": "career_readiness",
  "status": "loaded"
}

---

POST /predict

Accepts a student profile and returns three ML predictions.

Request Body

Field	Type	Constraints	Description
age	int	16–40	Student age
gender	str	—	Male, Female, Non-binary
degree_type	str	—	B.Tech/B.E., M.Tech/M.Sc, MBA, BCA/MCA
stream	str	—	CS/IT, Engineering (Non-CS), Commerce/Management
year_of_study	int	1–6	Current year of degree
college_tier	str	—	Tier 1, Tier 2, Tier 3
urban_or_rural	str	—	Urban, Rural
daily_ai_tool_usage_hrs	float	0–24	Hours per day using AI tools
primary_ai_tools_used	str	—	ChatGPT, Gemini, Claude, Perplexity, GitHub Copilot, None
ai_replaces_own_thinking_score	int	1–10	Self-rated AI reliance for cognitive tasks
ai_dependency_score	int	1–10	Overall AI dependency self-assessment
fear_of_job_loss_to_ai	int	1–10	Perceived threat from AI to career
career_clarity_score	int	1–10	Clarity about career direction
internship_experience	int	≥ 0	Number of internships completed
weekly_job_application_count	int	≥ 0	Job applications per week
resume_confidence_score	int	1–10	Confidence in resume quality
interview_anxiety_score	int	1–10	Anxiety level during interviews
daily_study_hours	float	0–24	Hours per day studying
self_learning_hours_per_week	float	≥ 0	Self-directed learning hours per week
skill_development_courses_taken	int	≥ 0	Online courses completed
social_media_hrs_per_day	float	0–24	Daily social media usage
sleep_hours	float	0–24	Average nightly sleep
stress_level	int	1–10	Self-rated stress
motivation_score	int	1–10	Self-rated motivation

Response

{
  "burnout_score": 6.24,
  "career_counseling_score": 1.0,
  "career_readiness_score": 7.83
}

Field	Range	Interpretation
burnout_score	1–10	Predicted burnout level (higher = more burned out)
career_counseling_score	0 or 1	Binary: 1 = likely to seek career counseling
career_readiness_score	continuous	Predicted overall career readiness (higher = more ready)

---

License

MIT © 2026 Soham Lodh. See LICENSE for details.