SentinelAI is a banking-grade fraud investigation platform designed to detect AI-generated phishing content, synthetic documents, and manipulated media across enterprise communication channels.
It combines machine learning detection, fraud-intent scoring, LLM-assisted explanations, and structured case management into a unified workflow system tailored for financial institutions.
AI-generated fraud is evolving faster than traditional rule-based filters can respond. Financial institutions now face:
- Synthetic phishing emails crafted by large language models
- AI-assisted payment redirection scams
- Hybrid attacks blending human-written and AI-generated content
- Deepfake-based identity manipulation
SentinelAI addresses this challenge by combining:
- AI-generation detection
- Fraud-intent classification
- Policy-based escalation logic
- Investigation case lifecycle management
- Compliance-ready audit logging
This is not just a detection engine — it is a fraud investigation assistant aligned with banking operations.
SentinelAI is designed for:
- Fraud investigation teams
- Corporate banking relationship managers
- Compliance officers
- Security operations teams
- Risk governance departments
Instead of manually evaluating suspicious content, users can:
- Submit content for analysis
- Receive AI and fraud probability scores
- Review highlighted suspicious signals
- Escalate through a structured workflow
- Generate investigation-ready reports
This reduces manual investigation time and improves decision confidence.
A corporate client forwards an urgent email requesting a change in payment instructions.
Traditional approach:
- Manual tone review
- Sender verification
- Escalation if unsure
With SentinelAI:
- Instant AI-generation and fraud scoring
- Highlighted suspicious phrases
- Structured risk classification
- Automatic case creation for policy-level risks
- Full investigation timeline tracking
This ensures consistent and auditable decision-making.
┌─────────────────────────────────────────────────────────────────┐ │ Next.js Frontend │ │ Dashboard │ Cases │ Analytics │ Clients │ Governance │ └────────────────────────────┬────────────────────────────────────┘ │ REST API ┌────────────────────────────┴────────────────────────────────────┐ │ FastAPI Backend │ │ │ │ Text │ Document │ Image │ Auth (JWT + RBAC) │ │ │ │ ML Inference Pipeline │ │ - AI Detector (sklearn) │ │ - Fraud Detector (sklearn) │ │ - Stylometric Analysis │ │ - Keyword Fraud Signals │ │ - Ollama Mistral (LLM Explanation) │ │ │ │ Risk Escalation Engine │ │ Case Management System │ │ PDF Report Generator │ │ │ │ SQLite Enterprise Schema │ │ audit_logs │ cases │ case_notes │ clients │ users │ └──────────────────────────────────────────────────────────────────┘
Text Channel:
- Stylometric feature extraction
- AI-generation classification
- Fraud-intent keyword scoring
- Risk fusion logic
Document Channel:
- PDF extraction
- Full text fraud pipeline
- Metadata inspection
Image Channel:
- Deepfake probability estimation
- Pixel-level statistical analysis
Risk formula:
risk_score = (0.6 × fraud_probability) + (0.4 × ai_probability)
Risk Levels:
| Level | Score Range | Recommended Action |
|---|---|---|
| LOW | 0.0 – 0.3 | Log only |
| MEDIUM | 0.3 – 0.6 | Manual review |
| HIGH | 0.6 – 0.8 | Strong review |
| CRITICAL | 0.8 – 1.0 | Immediate escalation |
Escalation types:
- critical_risk
- human_crafted_fraud
- synthetic_suspicious
- elevated_risk
OPEN → UNDER_REVIEW → ESCALATED → RESOLVED / FALSE_POSITIVE
Features:
- Auto-case creation for MEDIUM+ risk
- Investigator assignment
- Threaded notes with timestamps
- Timeline tracking
- One-click PDF export
- Link cases to enterprise clients
- Aggregate risk trends per client
- Track repeated fraud attempts
- View risk distribution summaries
Dashboard metrics include:
- Daily fraud detection trends
- AI-generated fraud percentage
- Risk distribution charts
- Average case resolution time
- Top fraud signals
- Case pipeline overview
- JWT authentication
- Role-Based Access Control (Analyst, Reviewer, Admin)
- Endpoint-level authorization
- SHA-256 content hashing
- Immutable audit trail
- Controlled case resolution permissions
Frontend:
- Next.js 14
- React 18
- TailwindCSS
- Recharts
Backend:
- FastAPI
- Python 3.11
- Async SQLite
ML:
- scikit-learn
- Stylometric feature extraction
Deepfake:
- PyTorch + heuristic analysis
LLM:
- Ollama (Mistral 7B, local inference)
Auth:
- python-jose (JWT)
- passlib (bcrypt)
Reporting:
- ReportLab
Database:
- SQLite enterprise schema
- Enron Email Dataset (corporate baseline)
- Phishing Email Dataset (fraud patterns)
- AI vs Human Text Dataset (AI detection)
- Deepfake datasets (image manipulation detection)
Training uses sampled subsets for real-time inference performance.
Backend:
cd backend pip install -r requirements.txt uvicorn app.main:app --reload
Frontend:
cd frontend npm install npm run dev
Ollama:
ollama pull mistral ollama serve
Content Submitted │ ▼ ML + Fraud Analysis │ ▼ Risk Evaluation │ LOW → Audit Log Only │ MEDIUM+ │ ▼ Auto-Create Case │ ▼ Escalation Logic │ ▼ Investigation & Resolution │ ▼ PDF Report Export
SentinelAI emphasizes:
- Decision support over blind automation
- Explainability over opacity
- Governance over novelty
- Workflow integration over isolated prediction
MIT License
Built as an enterprise-ready hackathon proof-of-concept.
EOF