RecruitShieldAI

Autonomous AI recruiter co-pilot powered by AWS & Gemini, featuring a 5-Point Anomaly Firewall to neutralize resume fraud, eliminate bias, and accurately rank top tech talent.

Comment

Inspiration

Modern technical hiring is broken by two converging crises: resume fraud and algorithmic bias. With LLM-assisted resume generators, bad actors flood job portals with synthetic "honeypot" profiles — inflating skill timelines and manipulating keywords to bypass legacy ATS filters. At the same time, naive ranking algorithms unfairly penalize high-potential candidates based on company background. We created RecruitShieldAI to solve both problems simultaneously: an autonomous recruiter co-pilot built with an integrated 5-Point Anomaly Firewall that neutralizes resume fraud, mitigates hiring bias, and empowers talent teams with transparent AI reasoning.

What it does

RecruitShieldAI is an end-to-end autonomous recruiting co-pilot:

  • 5-Point Anomaly Firewall: Automatically scans applicant pools to detect and quarantine synthetic trap profiles by identifying impossible skill durations, post-dated signups, and timeline contradictions.
  • Bias-Mitigated Scoring Engine: Replaces strict disqualifications with dynamic soft-scoring penalties (for example, soft score adjustments for IT consulting backgrounds to prevent unfair rejection while rewarding product impact).
  • Precision Match Ranking: Dynamically ranks candidates against job descriptions with transparent recruiter rationale and match scores.
  • AI Co-Pilot Chatbot: Powered by Gemini 2.5 Flash, recruiters can ask natural language questions like "Who is ranked #1?", "Who has React experience?", or "Explain honeypots" and receive instant, precise candidate summaries.
  • One-Click Shortlist Export: Easily export verified, screened top candidates directly into clean structured formats for seamless hiring team collaboration.

How we built it

RecruitShieldAI was built using a robust modern tech stack:

  • Frontend: React 18, Vite, TypeScript, TailwindCSS, and a glassmorphic dark-mode UI with live animated scanning metrics.
  • Backend Framework: Python FastAPI server hosted on Render with dataset support for JSONL and CSV candidate pools.
  • Agentic AI & LLMs:
    • AWS Bedrock & Strands Agents: Powers the multi-step screening pipeline for audit integrity, consulting filters, and candidate ranking.
    • Google Gemini 2.5 Flash (google-genai): Provides ultra-fast, context-aware recruiter Q&A over the candidate dataset.
    • Sentence Transformers (bge-base-en): Generates semantic embeddings for precise vector-based candidate matching.

Challenges we ran into

  • Synthetic Resume Detection: Establishing precise heuristic rules to catch sophisticated honeypot trap profiles without flagging legitimate candidates required rigorous cross-validation of work experience versus skill durations.
  • Dual-AI Orchestration: Seamlessly combining AWS Bedrock for structured agentic screening pipelines and Gemini 2.5 Flash for conversational co-pilot Q&A into a unified API.
  • Clean AI Synthesis: Preventing raw JSON tool outputs from leaking into the user chat UI and ensuring every query returns clean, structured markdown.

Accomplishments that we're proud of

  • High Anomaly Detection Accuracy: Successfully detected and quarantined synthetic trap profiles across candidate pools without a single false positive.
  • Instant Conversational AI: Built a recruiter chatbot that answers complex dataset questions in under 2 seconds using Gemini 2.5 Flash.
  • Responsive Glassmorphic UI: Designed an intuitive dashboard that gives recruiters clear visual feedback during every stage of candidate screening.

What we learned

  • Explainability is Essential: Recruiters trust AI only when every score adjustment, rank position, and security flag comes with clear, human-readable reasoning.
  • Soft Scoring Beats Hard Filters: Soft penalty adjustments preserve promising candidates who would otherwise be filtered out by traditional keyword ATS systems.

What's next for RecruitShieldAI

  • Multi-Tenant ATS Integrations: Direct integration plugins for Greenhouse, Lever, and Workday.
  • Blind Screening Mode: Zero-knowledge privacy layer for unbiased initial candidate evaluations.
  • Multimodal Video Screening: Automated AI analysis for asynchronous candidate video interviews.

Built With

Share this project:

Updates

posted an update —

Quick Pro-Tip for Judges & Testers! If loading the demo candidate dataset takes a moment during initial setup (due to cloud backend cold-starting), please perform a hard refresh (Ctrl + Shift + R on Windows/Linux or Cmd + Shift + R on Mac). This instantly clears cached assets and speeds up candidate data ingestion & AI co-pilot queries!

Log in or sign up for Devpost to join the conversation.

Submission history