RETAIN.AI | Employee Attrition Predictor

PLATFORM WALKTHROUGH

Strategic AI for Talent Retention & Cost Optimization

Business Impact

Employee turnover is a multi-billion dollar problem. This project provides HR teams with an Intelligence Report that predicts "Flight Risk" with ~70% accuracy, allowing for proactive intervention. By automating the data extraction from resumes and predicting attrition, this tool directly aims to reduce turnover costs and improve organizational stability.

• Core Implementation •

• • • •

• Languages •

Dataset

➤ Human Resources Data Set by Dr. Carla Patalano and Dr. Rich Huebner

Key Features

• GenAI Resume Parsing: Uses Google Gemini Pro to extract complex employee metrics from uploaded PDFs, bypassing manual data entry.
• Predictive Risk Modeling: Implements a cost-sensitive XGBoost classifier to identify high-risk employees.
• Dynamic HR Dashboard: A sleek, dark-themed UI featuring Risk Gauges, Intelligence Reports, and AI-generated Strategy Recommendations.
• Automated Intelligence Reports: Generates full-scale PDF/CSV detailed reports, enabling HR leads to transition from raw data to board-ready presentations instantly.
• Enterprise-Ready Data Flow: Secure handling of employee records using MongoDB Atlas, ensuring data persistence and historical trend tracking (MDE).

---

Tech Stack

Category	Technology	Implementation
Frontend	React (Vite), Tailwind	Responsive SPA with a dark-themed HR workspace.
Frontend	Fetch API	Utilizing native browser APIs for asynchronous data fetching and promise-based HTTP requests.
Backend	FastAPI	Building a high-performance, asynchronous REST API with automatic OpenAPI documentation.
Backend	Pydantic	Enforcing strict data validation and type-safe schemas for incoming employee data.
Parsing	PyMuPDF (fitz)	Low-level PDF binary stream extraction before GenAI processing.
Security	CORS Middleware	Orchestrated Cross-Origin Resource Sharing for secure Vercel-to-Render communication.
Database	Motor (Async MongoDB)	Non-blocking database drivers to ensure high-concurrency performance.
AI Engine	Google Gemini Pro	Leveraging Large Language Models (LLMs) for intelligent data extraction from PDF resumes.
ML Model	XGBoost	Deploying Gradient Boosted Decision Trees with cost-sensitive weights for risk classification.
Deployment	Vercel, Render	Distributed cloud hosting with automated CI/CD.

System Architecture

1. Data Acquisition (Dual Entry):
   • Automated Path: User uploads a CV via React; FastAPI orchestrates the file stream to Google Gemini Pro for entity extraction.
   • Manual Path: Users can directly input employee metrics into a structured form, bypassing the AI extraction for immediate results.
2. Standardization & Validation: Both data paths converge at the Pydantic layer, which enforces strict schema validation and type-safety before the data reaches the ML model.
3. Intelligence Layer (Inference): The validated data is processed by a pre-trained XGBoost model.
4. Cloud Persistence: Prediction logs and metadata are asynchronously committed to MongoDB Atlas using the Motor driver.
5. Real-Time Analytics: The dashboard leverages MongoDB Aggregation Pipelines ($match, $group, $avg) to offload heavy computations to the database. This enables real-time tracking of Risk Hotspots (high-risk departments) and Organizational Averages. Data is fetched via the native Fetch API for instant UI synchronization.

Challenges Faced During Development

1. Gemini API Rate-Limit Management

The Problem: During the integration of Google Gemini, the system frequently hit API rate limits despite low usage volume, and the frontend-to-backend data stream was failing to trigger the extraction logic correctly.

• The Pivot: - Model Switching: Swapped models to optimize token usage and cost-efficiency.
• Robust Debugging: Implemented a comprehensive logging and "Safety Mechanism" layer to catch rate-limit exceptions before they crashed the frontend.

2. Prioritizing Business Impact (From SMOTE to Cost-Sensitivity)

The Problem: Initial attempts to handle dataset imbalance using SMOTE (Synthetic Minority Over-sampling Technique) resulted in lower Precision and Recall, as the synthetic data introduced noise that hindered the model's ability to generalize to real employee behavior.

• The Pivot: I pivoted to Cost-Sensitive Learning by tuning the scale_pos_weight parameter & I abandoned the resampled X_train_res dataset in favor of the original, authentic X_train. This forced the model to learn from real-world distributions while penalizing the misclassification of flight risks, significantly improving the model's predictive reliability.

Metric	Old Model (SMOTE)	New Model (Weighted XGBoost)	Impact
Overall Accuracy	67.00%	69.84%	+2.84% Improvement
Precision (Class 1)	53.00%	60.00%	Higher reliability in flags
Recall (Class 1)	41.00%	41.00%	Consistent detection rate
Class 0 Recall	80.00%	85.00%	5% fewer False Positives
Data Integrity	Synthetic	Organic (Original)	No "hallucinated" data

Local Development Setup

Backend (FastAPI)

1. Navigate to /backend and create a .env file with your GEMINI_API_KEY and MONGO_URI.
2. Install dependencies: pip install -r requirements.txt
3. Run the server: uvicorn main:app --reload

Frontend (React)

1. Navigate to /frontend and install dependencies: npm install
2. Start the development server: npm run dev

References

SMOTE • No module error • Deprecated dict in Pydantic • Gemini API quickstart • MongoDB Atlas pipeline stages

---

Made by Neha K Vallappil • LinkedIn