README.md

metadata

title: AI Recruiting Agent
emoji: 🤖
colorFrom: indigo
colorTo: blue
sdk: gradio
python_version: '3.10'
app_file: app.py
fullWidth: true
header: default
short_description: Bias-aware RAG AI recruiting agent with hallucination checks
suggested_hardware: cpu-upgrade
pinned: true
tags:
  - gradio
  - langchain
  - chroma
  - rag
  - recruiting
  - hr-tech
  - responsible-ai
  - fairness
  - bias-mitigation
  - llm
  - vector-database
  - ai-agent
  - nlp
  - demo
  - explainable-ai
  - enterprise-ai
models:
  - openai/gpt-oss-120b
thumbnail: >-
  https://huggingface.co/spaces/19arjun89/AI_Recruiting_Agent/raw/main/thumbnail.png
disable_embedding: false
startup_duration_timeout: 45m

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

🧠 AI Recruiting Agent

A transparent, verification-driven, and bias-aware recruiting assistant built with Gradio, LangChain, and vector search.

This project is designed as a decision-support system to help recruiters evaluate candidates and generate outreach — while embedding safeguards for fairness, auditability, and explainability.

📘 For a detailed technical and governance walkthrough, see:
Guidebook.md

---

🚀 What This Tool Does

1) Candidate Assessment (Recruiter View)

Recruiters can:

• Upload company culture documents (core values, principles, etc.)
• Upload resumes in bulk
• Paste a job description

For each candidate, the system produces:

• Deterministic technical skills score
• Evidence-backed culture fit analysis
• Policy-constrained hiring recommendation
• Automated fact verification
• Structured bias audit

All outputs are designed to be inspectable and reviewable by humans.

---

2) Cold Email Generator (Candidate View)

Candidates can:

• Upload a resume
• Paste a job description

The system generates a tailored professional outreach email.

This feature is independent of the assessment pipeline.

---

✨ Key Differentiators

Most recruiting tools focus on automation and ranking.

This system prioritizes trust and governance:

• ✅ Evidence-first scoring
• ✅ Self-verifying analyses
• ✅ Deterministic policies
• ✅ Bias and fairness auditing
• ✅ Human-in-the-loop controls

It is built to support high-stakes hiring environments, not replace recruiters.

---

🛡️ Responsible AI Safeguards

This prototype embeds multiple layers of protection against hallucination and bias.

🔹 Input Anonymization

Before indexing:

• Emails, phone numbers, URLs, addresses, and demographics are removed
• Likely name headers are masked

This reduces demographic leakage into retrieval and reasoning.

---

🔹 Deterministic Verification

All required skills are:

• Extracted by LLM
• Verified against the job description
• Classified as:
  • Quote-verified
  • Name-only
  • Dropped (unsupported)

Only quote-verified skills are scored.

---

🔹 Evidence-Based Fact Checking

All claims in:

• Skills analysis
• Culture analysis

Are verified against:

• Resume text
• Job description
• Culture documents

Unsupported claims are flagged and may trigger automatic revision.

---

🔹 Bias Audit Chain

Each candidate run includes a bias audit that inspects:

• Pedigree and prestige bias
• Subjective or exclusionary language
• Inconsistent standards
• Non-job-related reasoning
• Overconfident claims

Outputs include:

• Bias indicators
• Recruiter guidance

These signals never auto-disqualify candidates.

They highlight where human review is essential.

---

📘 System Design & Governance

This project is documented in detail in:

➡️ guidebook.md

The guidebook covers:

• End-to-end system flow
• Deterministic vs LLM boundaries
• Scoring rules
• Verification logic
• Implied competency handling
• Self-correction policies
• Audit artifacts
• Model governance

Reviewers, auditors, and contributors are encouraged to consult it.

---

🧪 Try It Out

Candidate Assessment

1. Upload culture documents (PDF/DOCX/Text)
2. Upload resumes (PDF/DOCX/Text)
3. Paste job description
4. Click Analyze Candidates

Cold Email Generator

1. Upload resume
2. Paste job description
3. Click Generate Cold Email

Sample documents are available in the repository.

---

⚠️ Human-in-the-Loop Disclaimer

This tool provides structured decision support.

It does not replace:

• Recruiter judgment
• Legal review
• Hiring policy compliance

Final hiring decisions must always be made by humans.

---

🧱 Technical Stack

• Gradio — UI
• LangChain — LLM orchestration
• Chroma — Vector storage
• ChatGroq — Inference
• Hugging Face Embeddings — Semantic search

---

📣 Feedback Welcome

This is an evolving research prototype.

Feedback is welcome on:

• Bias mitigation
• Evaluation transparency
• UX
• Failure modes
• Responsible AI patterns

Please share feedback via the Community tab.

---

Built by Arjun Singh