Guardian Recruit: Hybrid Fraud Detection & XAI for Digital Recruitment

Team Lead · System Architecture, Fusion Layer & Explainability | UNT DTSC 5082

February 2026 – May 2026 · 4-person capstone team

Project Summary

Guardian Recruit is my master’s capstone: a decision-ready fraud detection system for digital recruitment platforms. Three independent model streams feed into a single interpretable fraud score — built as a team, with each member owning a stream.

Team: Hemanth Kumar Gunda (NLP Stream — BERT fine-tuning), Srijitha Ungarala (NLP Stream — preprocessing & linguistic EDA), Kusuma Satya Sreeja Chalasani (Outlier Stream — IsolationForest), and me as team lead.

My role was system architecture and the parts that turn two separate model streams into one decision-ready product: the Fusion Layer, the explainability layer, the data/augmentation pipeline, and deployment. I did not build the BERT or IsolationForest streams myself — those were my teammates’ work — but I designed how they combine and how a human ends up trusting the output.

🔑 Key Contributions

1. Fusion Layer & System Architecture

Designed and built the XGBoost meta-classifier that fuses the BERT NLP stream’s fraud probability with the IsolationForest stream’s anomaly score into a single fraud_score, plus adversarial meta-features (domain age via WHOIS, GPT-2 text perplexity, messaging-platform risk). Result: 0.9718 ROC-AUC, 99% accuracy, 0.8439 fraud F1 on a held-out validation set of 1,877 postings (30 false negatives, 7 false positives).

2. Explainable AI (XAI)

Built the SHAP-driven explainability layer (shap_explainer.py) that produces a per-prediction reasoning summary for every verdict — e.g. “Flagged because: NLP fraud pattern score (+86%), No company logo (+7%), Outlier metadata pattern (+4%)” — plus a ChromaDB-backed RAG explainer that retrieves similar known fraud cases and an LLM narrative layer (Groq → Ollama → template fallback).

3. Data Engineering & Adversarial Robustness

Owned the data pipeline: ingesting the EMSCAD dataset (17,880 postings, 4.85% fraud rate), generating 200 synthetic 2026-era fraud listings grounded in FBI IC3, FTC, and BBB reports (crypto salaries, Telegram interviews, equipment-deposit scams), and applying SMOTENC to expand the training set to 17,342 rows.

4. Deployment & Fairness Auditing

Deployed the live Streamlit demo on Hugging Face Spaces and ran a subgroup fairness audit across geography, employment type, and industry — surfacing real risks like the model under-detecting contract-style fraud and over-flagging logo-less small businesses, and documenting mitigation steps (SHAP transparency, independent keyword signals, RAG case comparison).

🔍 See It Work

A live posting scored fraud_score = 0.94 with the reasoning: “NLP fraud pattern score (+86%), No company logo (+7%), Outlier metadata pattern (+4%), No screening questions (+2%)” — triggered keyword signals included compensation_guarantee, messaging_app_interview, and equipment_bait. Try the live demo →