Kritikos AI is an end-to-end human-in-the-loop AI decisioning system for document-driven workflows such as insurance claims, underwriting, fraud review, and compliance checks.
The system demonstrates how AI assists human decision-makers rather than replacing them — a pattern commonly used in high-stakes, real-world platforms.
Experience the Kritikos AI UI flow without running the code locally:
- Customers upload documents (PDFs)
- AI extracts structured data and computes a risk score
- A unique reference ID is generated per submission
- All outputs are persisted in a database
- Investigators retrieve claims using the reference ID
- Investigators review AI output (risk score + explanation)
- Investigators can approve, reject, or override the AI decision
- Final decisions are stored and auditable
Most AI demos stop at model output.
Real-world systems require:
- traceability
- human oversight
- explainability
- persistence
- workflow handoffs across roles
Kritikos-AI focuses on end-to-end system design, not just AI scoring logic.
Customer
↓
FastAPI Backend
↓
AI Scoring Engine
↓
SQLite (decisions table)
↓
Investigator Review & Override
- Reference ID acts as the single contract between Customer and Investigator
- AI recommendations are advisory, not final
- Humans retain decision authority
- All decisions and overrides are persisted
- Backend: FastAPI (Python)
- Database: SQLite (local persistence)
- AI Logic: Rule-based, configurable scoring engine
- Frontend:
- Customer UI (document upload)
- Investigator UI (claim review & decision)
- Version Control: Git + GitHub
POST /submit-claim/
- Uploads documents
- Triggers AI scoring
- Returns a reference_id
GET /claims/{reference_id}
- Retrieves claim details
- Includes AI risk score, explanation, and status
POST /claims/{reference_id}/decision
- Approve, reject, or override AI decision
- Add investigator comments and escalation details
- Customer upload → AI scoring → database persistence → investigator review → human override
- Fully tested locally using curl and UI flows
- Backend hardened against schema drift and file-path issues
- Authentication and role-based access are intentionally out of scope
- Focus is on decision workflow correctness, not production hardening
- Designed as a portfolio-quality system prototype
- Authentication and role-based access
- Investigator dashboard with claim listing
- Decision audit trail with timestamps
- ML-based scoring models
- Cloud-hosted database integration
Built by Sai Anurag Balijepalli
GitHub: https://github.com/AiPM-Anurag
This project is intended for learning, demonstration, and portfolio purposes.