Building an AI-Powered CV Screening System with Explainable Scoring

Website Demo : https://hireco.intelix.fun/
Github : https://github.com/nesnyx/Hireco-AI

Currently i joined Hackthon and i build this app.

Building an AI Scoring System for Smarter and More Transparent CV Screening

Recruitment can be exhausting — hundreds of CVs come in, and HR teams have to manually scan each one just to find a few strong candidates.
As a developer fascinated by how AI can augment human decision-making, I asked myself:

“Can an AI objectively evaluate CVs based on job criteria — and still explain its reasoning transparently?”

That question became the foundation for my latest project:
🎯 AI Scoring CV Screening System — an intelligent engine that can:

Analyze CVs (PDF/Text)

• Score candidates against custom job criteria
• Explain the reasoning behind each score
• Store results in a structured database
• Allow candidate comparison within the same job posting

System Architecture Overview

The system is designed with modular and scalable components:

1. Frontend (Reactjs Vite) — user interface for uploading CVs and viewing AI results
2. Backend (FastAPI) — handles upload, analysis pipeline, and LLM integration
3. LLM Engine (GeminiAI / Local Model) — interprets CVs contextually based on job descriptions
4. Vector Database (ChromaDB) — stores semantic embeddings for similarity search and clustering
5. Relational Database (SQLite / Postgres) — keeps structured analysis results and scoring logs

🔍 Before sending data to the LLM, an embedding-based pre-filter reduces unnecessary API calls — saving both cost and time.

Top 3 Improvement Priorities:

1. Add customizable scoring criteria (e.g., weight for skills, experience, presentation).
2. Enable candidate comparison dashboards per job posting.
3. Implement bias and fairness monitoring for ethical scoring.

💡 Lessons Learned

1. AI is not just about accuracy — it’s about interpretability.
2. HR users value why a score was given, not just the number.
3. Efficiency matters.
4. Embedding pre-filters cut inference costs by up to 70%.
5. Deployment gotchas teach real lessons.
6. The infamous attempt to write a readonly database error taught me to handle Docker file permissions carefully.
7. Human + AI collaboration works best.
8. The goal isn’t replacing HR — it’s empowering them with better insights.

🛠️ Tech Stack

• Backend: FastAPI + Celery (for async jobs)
• AI Engine: LangChain / LLM (Gemini or local model)
• Databases: SQLite/PostgreSQL + ChromaDB
• Deployment: Docker Compose

🌍 Next Steps: Toward an HR-AI Platform

• The roadmap ahead focuses on scaling and usability:
• Add criteria templates for HR teams
• Provide candidate comparison visualization
• Offer LLM explainability reports for audit transparency
• Integrate with job portals via API

🧩 Reflection

This project isn’t just about automation — it’s about rethinking how AI can assist decision-making.
An explainable scoring system gives HR teams confidence that AI isn’t a “black box,” but a partner in fair evaluation.

🧠 AI should not replace human judgment — it should strengthen it.

💬 Let’s Connect
If you’re working on:

AI explainability
LLM applications in HR or recruitment
Human-AI collaboration design