Self-Hosted NLP Service Documentation

Beginners
1

Help me go through this short documentation, and confirm if my self -hosted nlp service is best for cost-reduction. And will I get best result from it?

Self-Hosted NLP Service Documentation

Overview

This documentation explains the comprehensive self-hosted Natural Language Processing (NLP) service implemented for the Wasefumi job application platform. The service replaces expensive OpenAI API calls with cost-effective, privacy-focused local processing while maintaining all existing functionality.

:bullseye: Objectives

  • Cost Reduction: Eliminate 90%+ of AI-related costs by replacing OpenAI API calls

  • Privacy Enhancement: Keep sensitive resume and job data on-premises

  • Performance Optimization: Reduce latency by eliminating external API dependencies

  • Feature Parity: Maintain all existing ATS analysis, job matching, and cover letter generation features

:building_construction: Architecture Overview


┌─────────────────┐ ┌──────────────────┐ ┌─────────────────┐

│ Frontend │ │ Backend API │ │ NLP Service │

│ (Next.js) │───▶│ (Express.js) │───▶│ (Self-hosted) │

└─────────────────┘ └──────────────────┘ └─────────────────┘

│ │

▼ ▼

┌──────────────────┐ ┌─────────────────┐

│ Supabase DB │ │ ML Models │

│ (PostgreSQL) │ │ (Transformers) │

└──────────────────┘ └─────────────────┘

:hammer_and_wrench: Technology Stack

Core NLP Libraries

  • @xenova/transformers: Lightweight transformer models for browser/Node.js

  • natural: Classic NLP algorithms and utilities

  • compromise: Natural language understanding and parsing

  • stemmer: Word stemming for keyword normalization

  • stopword: Remove common words for better keyword extraction

Machine Learning Models

  • DistilBERT: Text classification and sentiment analysis

  • MiniLM-L6-v2: Sentence embeddings for semantic similarity

  • BERT-NER: Named entity recognition for resume parsing

Backend Integration

  • Node.js/TypeScript: Runtime and type safety

  • Express.js: API routing and middleware

  • Supabase: Database and authentication

  • Joi: Input validation

1 Like
2

I don’t think there are any apparent flaws, especially in terms of cost.

1 Like
3

what would you advice. should I proceed with it

1 Like
4

Hmm… For example, if the scale is small (e.g., less than 100K requests per month), it may be more cost-effective to use an existing well-known AI API.
Whether you use your own hardware or rent it, it won’t be free…

If you decide to proceed, I don’t think there will be any technical issues with the plan itself, but if you are unsure about implementing each part, you could look for existing open source projects that are similar and combine them.

https://github.com/topics/applicant-tracking-system

5

a well-known AI API like