Multimodal OCR & Single-Pass Translation for Foreign Receipts
How SplitBill parses non-Latin scripts (Thai, Georgian, Chinese), reconciles line items to the penny, and eliminates hallucinations with a three-layer defense.
Single-Pass Vision-Language Inference
Traveling groups often encounter receipts in non-Latin scripts where item names like ต้มยำ or ხაჭაპური are unintelligible. SplitBill passes the raw receipt photo through a multimodal Vision-Language Model (Qwen2.5-VL or Gemini Flash) that extracts structured JSON line items and translates dish names into the user's language in a single inference pass.
Three-Layer Anti-Hallucination Defense
LLMs can make arithmetic mistakes. SplitBill prevents hallucinations through a 3-layer pipeline: image preprocessing (auto-rotation, CLAHE contrast enhancement), constrained decoding strictly enforcing a JSON Schema, and deterministic arithmetic reconciliation. If the sum of extracted items plus taxes and service charges deviates from the receipt total by more than 0.05, the system automatically flags the receipt for manual review rather than propagating bad math.
Zero Marginal Cost & Local Privacy
By supporting local execution via Ollama, receipts are processed entirely on private hardware without per-scan API costs. Personal purchase histories and merchant locations stay confidential, and fallback to cloud vision models remains completely transparent and optional.