Skip to content

ml-service.md

Latest commit

 

History

History
225 lines (146 loc) · 4.4 KB

ml-service.md

File metadata and controls

225 lines (146 loc) · 4.4 KB

ML Service – Resume Analyzer & Moderation API

This microservice is responsible for all machine learning capabilities in the platform, including resume analysis, job recommendations, and real-time message moderation.

It is built using FastAPI and deployed as an independent service to ensure scalability and modularity.

Overview

The ML service processes user inputs such as resumes and chat messages, applies trained machine learning models, and returns structured insights used across the application.

Responsibilities

  • Resume parsing and analysis
  • Job recommendation based on dataset similarity
  • Skill extraction and gap detection
  • Toxic message detection
  • Spam message classification

Core Features

Resume Analysis

  • TF-IDF based vectorization of resume and job description
  • Cosine similarity scoring for compatibility
  • Extraction of relevant skills from resume text
  • Identification of missing skills compared to job description
  • Generation of structured analysis output

Job Recommendation System

  • Uses a preprocessed job dataset
  • Matches resumes with similar job roles
  • Provides relevant job suggestions
  • Integrates external job APIs (Adzuna) for live listings

Message Moderation System

Spam Detection

  • Model: Naive Bayes
  • Vectorization: TF-IDF
  • Purpose: Detect unwanted or promotional messages

Toxicity Detection

  • Model: Logistic Regression
  • Purpose: Identify harmful, abusive, or inappropriate content

Model Architecture

Models Used

  • job_dataset.pkl → Preprocessed dataset for job matching
  • toxic_model.pkl → Toxicity classification model
  • spam_model.pkl → Spam classification model
  • tfidf_vectorizer copy.pkl → Vectorizer for resume analysis
  • spam_vectorizer.pkl → Vectorizer for spam detection

Model Hosting Strategy

Due to GitHub file size limitations, trained models are not stored in the repository.

Instead, they are hosted on Hugging Face Hub and downloaded dynamically during runtime.

Benefits

  • Keeps repository lightweight
  • Avoids large file issues (>100MB)
  • Enables easy model updates without redeploying backend
  • Supports scalable deployment

Model Loading Mechanism

Models are fetched at runtime using a helper function:

download_file(HF_URL, "filename.pkl")

Key Notes

  • Ensure correct Hugging Face URL format:
    • Use /resolve/main/
    • Do NOT use /blob/
  • Encode spaces in filenames:
    • Example: tfidf_vectorizer%20copy.pkl

Message Moderation

POST /moderate

Request:

{
  "text": "User message"
}

Response includes:

  • Spam classification
  • Toxicity classification

Deployment Configuration

Platform

Render (Cloud Deployment)

Settings

Root Directory: ml_services

Build Command

pip install -r requirements.txt

Start Command

uvicorn app:app --host 0.0.0.0 --port 10000

Local Development

Run the service locally:

uvicorn app:app --reload --port 80004

Environment Variables

Sensitive keys should not be hardcoded.

Use environment variables:

ADZUNA_APP_ID ADZUNA_APP_KEY

ML Pipeline

  1. Data collection and preprocessing
  2. Text cleaning and normalization
  3. Feature extraction using TF-IDF
  4. Model training and evaluation
  5. Model serialization using pickle (.pkl)
  6. Deployment via FastAPI
  7. Runtime inference through API endpoints

System Architecture

System Architecture

Design Decisions

Microservice Architecture

The ML logic is separated from the main backend to:

  • Improve scalability
  • Allow independent deployment
  • Reduce backend complexity
  • External Model Hosting

Using Hugging Face allows:

  • Faster deployments
  • Centralized model management
  • Easy version control for models

Limitations

  • Initial request may be slower due to model download
  • No caching implemented yet
  • Basic models (can be improved with deep learning)

Future Improvements

  • Lazy loading for faster startup
  • Redis caching for predictions
  • Docker containerization
  • Model versioning system
  • Rate limiting and authentication
  • Advanced NLP models (transformers)

Status

  • Production-ready
  • Fully deployed
  • Integrated with frontend and backend
  • Actively maintained