Hacki
About the project
KHIPU is an AI-powered invoice guardian for LATAM SMEs. It prevents duplicate payments, fake suppliers, and quantity mismatches by reconciling invoices against Purchase Orders (PO) and Delivery Guides before authorizing payment.
How it works:
- Local AI Extraction (QVAC): We use QVAC SDK with the QWEN3VL-2B multimodal model (Q4_K quantization, ~2.5GB) to extract structured data from messy invoice photos directly on the user's laptop. No data leaves the device.
- Deterministic Reconciliation: Our engine runs 8 strict checks (supplier match, quantity, price tolerance, math integrity, duplicate detection, delivery match, and unauthorized items).
- Traffic Light Decision: GREEN (auto-approve), YELLOW (human review for unauthorized items like unapproved freight), RED (block payment for critical errors).
- Web3 Payment (WDK): Approved invoices generate a secure USDT payment payload via Tether's WDK, ready for MetaMask execution on Sepolia.
Tech Stack: FastAPI, React, QVAC SDK (QWEN3VL-2B), Tether WDK, MetaMask. Transparency: The system is fail-closed. If the local model cannot run, it refuses to hallucinate and uses a verified ground-truth fallback or rejects the operation.